A0540
Title: How to measure relatedness between datasets during external validation of a multivariable prediction model
Authors: Harald Heinzl - Medical University of Vienna (Austria) [presenting]
Harbajan Chadha-Boreham - Clinical Biostatistics Consultancy (France)
Martina Mittlboeck - Medical University of Vienna (Austria)
Abstract: The purpose of a multivariable prediction model (MPM) in clinical practice is the diagnosis or prognosis of a disease. Before routine use, the performance of the MPM has to be evaluated via internal and external validation. The main task of external validation is the assessment of the MPM's generalisability, an umbrella term for reproducibility and transportability. If the development and validation population are closely related, then an external validation study assesses reproducibility, otherwise, it assesses transportability. Hence, relatedness measures between the development and validation population have to be defined and appropriately estimated from the corresponding datasets. Since the development dataset will usually not be available during external validation, these measures have to be based on those basic summary statistics (e.g. percentages, means and standard deviations) that are commonly reported in medical research papers. Three proposals for such measures will be presented. They will be exemplified by an external validation study of the Framingham steatosis index.
