Towards guidance on practical issues in multivariable model- building: The STRengthening Analytical Thinking for Observational Studies (STRATOS) initiative

• Applicant: Professor Dr. Wilhelm Sauerbrei
• Subject Area: Epidemiology and Medical Biometry/Statistics
• Term: since 2017
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 386466116

Design and analysis of observational studies often exhibit serious weaknesses. Lack of guidance on practical issues discourages many applied researchers from using more appropriate methods when analyzing observational studies. Furthermore, many analyses are conducted by researchers with limited experience in statistical methodology and software. Consequently, even standard analyses reported in the medical literature are often flawed, casting doubt on their results and conclusions. An efficient way to help researchers keep up with recent methodological developments is to develop guidance and suitable educational material. These observations led to the initiation of the STRengthening Analytical Thinking for Observational Studies (STRATOS) initiative, launched in August 2013, of which I am the chair. STRATOS is a collaboration of experts in different areas of biostatistical research, with over 100 members from 20 countries and all continents. STRATOS has nine topic groups and twelve cross-cutting panels. This project focuses on issues of topic group 2 (TG2) ‘Selection of variables and functional forms in multivariable analysis’, of which I am one of the co-chairs. In a TG2 overview paper we identified seven issues needing further research. Issues center around strategies for variable selection, the role of shrinkage and determination of the functional form for continuous variables. Edwin Kipruto and I gained further insight into the first two issues, e.g. the non-negative garotte can be used in data with stronger multicollinearity and in high-dimensional data, and published the protocol of a large simulation study to compare approaches for variable selection, with and without shrinkage. We also worked on a new R package for the multivariable fractional polynomial (MFP) approach and some of its extensions. Two MFP tutorial papers are planned, discussing issues like the required sample size and the role of influential points. Jointly with other TG2 members we will work on investigations of splines in a multivariable setting and comparisons to MFP in examples. TG2 reviewed a series of statistical (short) papers and tutorials published by medical journals with respect to various aspects of multivariable modelling. We identified relevant gaps and aim to improve educational material. We have started to work on Shiny apps and short videos. Well planned and carefully conducted simulation studies are the key instrument to investigate the properties of statistical approaches and comparing methods. To improve simulation studies, the ADEMP (aims, data-generating mechanisms, estimands, methods, and performance measures) concept was developed. To allow a fair comparison of statistical methods the concept of neutral comparison studies was proposed. Members of the STRATOS simulation panel will work on combining and extending the two concepts. For some methods from low-dimensional data the usefulness for analyses of high-dimensional data will be investigated.

• DFG Programme: Research Grants