Article
Variable selection and model instability [abstract of the session]
Search Medline for
Authors
Published: | February 26, 2021 |
---|
Outline
Text
Statistical modeling applying data-driven variable selection is a key aspect in the analysis of observational data. Most participants of the conference regularly use these methods. However, the resulting model instability is very often not acknowledged and not adequately dealt with.
Christine Wallisch will start the session with a short talk, presenting some case studies highlighting the problem of unstable statistical models. Afterwards Ewout Steyerberg will discuss his view on the issue of variable selection and model instability. He has been interested in this topic for many years. Riccardo de Bin will present an investigation of penalized regression methods on the effects of single observations on the optimal choice of the tuning parameter, the penalty. The penalty is often computed via cross-validation and its value heavily influences regression coefficient estimates and predictions. Finally, Georg Heinze will talk about some recent work on the consistency and accuracy of statistical measures quantifying model instability. He will also present a proposal on how to adequately present statistical models derived by data-driven variable selection (based on the R package ABE).
The aim of the session is to raise awareness of this important problem of statistical modeling (especially for younger or more applied researchers). We will present some answers how to solve the issue based on extensive simulation studies. After the session, attendees should know of principles and practical approaches to cope with model instability induced by variable selection.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.