Statistical Model Building Strategies for Cardiologic Applications (SAMBA)

点击翻译按钮获取中文摘要

Statistical models that adequately describe disease progression and treatment response are essential for development, improvement and judgment of therapies in all medical fields. The aims of such models are- to compare the outcome of interest between different interventions or levels of a modifiable risk factor while adjusting for confounding variables, - to identify explanatory variables and concisely describe the association between explanatory variables and the outcome of interest, and - to make predictions for the outcome of interest. The methodological needs and challenges for these three distinct aims are different. In medical applications, researchers often do not focus on a single aim but the interest lies in a combination. Generally, the development of a valid descriptive or explanatory model relies on the identification of a meaningfully sized set of explanatory variables and the specification of adequate functional forms. Intensive statistical research on both aspects was performed for decades. However, the results of this research are only poorly incorporated into clinical research. This interdisciplinary project intends to build a bridge between statistical research on model building strategies and implementation of the methodology into actual medical research. Therefore, this project aims at1. identifying deficiencies in multivariable models that were developed for cardiologic applications with respect to statistical model building (selection of variables and functional forms), 2. building real advanced statistical models for four typical cardiologic research questions by applying state-of-the-art methodology, 3. developing and evaluating new methods to correct for overestimation bias arising in data-driven model building, and4. providing guidance for model building strategies which are understandable for applied researchersTherefore, from a statistical point of view, the aim is to identify, discuss and improve the current standards applied in clinical research with respect to model building and variable identification. In this project, we focus on descriptive models, potentially combined with the aim to make a best possible prediction under possible sample size limitations. By this, the aim of the project lies in models where regression coefficients must be interpretable from a medical point of view. We will particularly address the impact of sample size and highlight options and limitations as occurring in real life situations. From a medical point of view, the aim is to gain new medical insights from statistical models which are built by employing better methodology. As a comprehensive result, we expect to be able to deduce methodologically improved and valid model building strategies for each of the four exemplary applications. We will use several original data sources of cardiovascular studies and combine them with results from the corresponding medical literature.