INTERACTIONS AND INDEPENDENCE IN STATED PREFERENCE MODELLING

In this paper two methodological problems associated to the estimation of discrete choice models using stated preference (SP) data in depth are examined. The first one is the incorporation of attribute interactions; this issue has largely been ignored (on the grounds that these are probably second order effects) because it implies a much larger data collection effort than if only the direct effects of the variables are considered. The second one is the effect of relaxing the hypothesis of independence between observations corresponding to a same individual when estimating SP models. To tackle the first problem an SP experiment for choice between car and car pool by students travelling to the main campus of the University was designed. Interactions are indeed important and it is shown that their best specification is not a trivial matter and that the extra burden on data collection has been largely overrated. To examine the second problem several data sets were used to test a variety of approaches proposed in the literature to deal with the problem. It was found that re-sampling methods do not always yield consistent results; it was also found that more exact methods seem to indicate that ignoring this problem may cause not only unduly large indices of precision of the estimates (t-ratios) but more seriously, bias in the values of the estimated parameters themselves. For the covering abstract, see IRRD E101013.