When a measured variable used to diagnose model behavior does not accurately represent the true underlying cause of errors.