A calibration problem in statistical diagnosis: The system transfer problem

Abstract

A statistical diagnostic system devised for a particular clinic may require appreciable modification when either the method of measurement of any of the diagnostic features changes within the clinic, or the system is to be applied to another clinic where different methods of measurement are used. In either circumstance the transfer can be effectively made only on the basis of information from a calibration experiment designed to establish the relationship between the different methods of measurement. Even with such calibration information a naive method of calibration commonly adopted is shown to yield conclusions which can be radically different from those of a method taking full account of the calibration unreliability. An application to a particular problem of clinical diagnosis is used to illustrate the analysis.