Abstract

An instrument that has been calibrated against a criterion measure with a sample of subjects is sometimes checked against the criterion in a validity study with another sample. In a spreadsheet-based simulation of such calibration and validity studies, a Bland-Altman plot of difference vs mean values for the instrument and criterion shows a systematic proportional bias in the instrument's readings, even though none is present. This artifactual bias arises in a Bland-Altman plot of any measures with substantial random error. In contrast, a regression analysis of the criterion vs the instrument shows no bias. The regression analysis also provides complete statistics for recalibrating the instrument, if bias develops or if random error changes since the last calibration. The Bland-Altman analysis of validity should therefore be abandoned in favor of regression. Verf.-Referat