A0821
Title: Comparing methods for flexible estimation of non-linear associations: The importance of suitable performance measures
Authors: Theresa Ullmann - Medical University of Vienna (Austria) [presenting]
Abstract: Flexible regression techniques (e.g., spline-based approaches or fractional polynomials) allow for modeling non-linear associations between continuous predictors and outcomes, often leading to improved model performance compared to specifying the associations as linear. Simulation studies are a key tool for comparing such techniques, but their conclusions critically depend on the performance measures used to evaluate how well the estimated curves recover the true underlying functions. A systematic categorization of performance measures is presented for evaluating estimated non-linear associations between an outcome and continuous predictors. To illustrate the practical implications of these choices, examples are presented that highlight how different performance measures can favor different methods. Results are also shown from a simulation study comparing several flexible modeling approaches, using performance measures from the proposed categorization. The emphasis is on the importance of aligning performance measures with the aim and the encouragement of more transparent and thoughtful evaluation strategies in methodological research.
