Abstract
IT IS ARGUED THAT THE STATISTICAL BASIS OF THE ANALYSIS OF SP DATA IS RATHER UNDER-RESEARCHED. WHEN SP DATA IS USED TO PROVIDE RELATIVE EVALUATIONS OF ATTRIBUTES, THE IMPLICATIONS ARE PROBABLY NOT SERIOUS: HOWEVER, THE FORECASTING ISSUES ARE MORE IMPORTANT. THESE ISSUES, RELATING TO MODEL SPECIFICATION AND DIFFERENT SOURCES OF ERROR, ARE MADE EXPLICIT IN RELATION TO THE RANDOM UTILITY THEORY COMMONLY USED WITH REVEALED PREFERENCE MODELS. IT IS CONCLUDED THAT THE USE OF MORE COMPLEX ESTIMATION METHODS SHOULD BE ENCOURAGED, AND THAT THERE REMAINS A NEED FOR 'EXTERNAL' DATA ON REAL CHOICES IF CREDIBLE FORECASTS ARE TO BE PRODUCED.(Author/publisher).