Stated Choice Experiments with Repeated Observations

Discrete choice models are an important tool in the analysis of travel behaviour (Ortuzar and Willumsen, 1990). The estimation theory of these models is mainly based on large sample results. This means that the number of observations, N9 on which estimations are based, should tend to infinity for the results to have desirable properties (such as, consistency). Empirical researchers sometimes worry about too small a value of N and try to increase it. In the context of stated choice modelling (an approach often used in travel demand modelling; see Bates, 1988) one may proceed in two directions: increase the number of respondents, or increase the number of observations per respondent. The latter alternative is usually much cheaper than the former. However, it gives rise to specific interdepend ency problems in the error structure (auto-correlation) which are ignored in the standard logit or probit models usually applied in this context (see Bates, 1988). The problem is similar to the auto-correlation problem encountered in the panel dataliterature. This paper shows how the solution from the panel data field (see, for example, Hsiao, 1986; Chamberlain, 1984) can be applied in the context of stated choice experiments. The method itself is not new. As far as we know, however, it has seldom (or never) been used in discrete choice models. It should, nevertheless, be realised that auto-correlation can be a serious problem. The purpose of this paper is therefore twofold: first, to demonstrate the application of the correct estimation method in the case of repeated observations on discrete choice; and second, to investigate the seriousness of the auto correlation problem. The model will be applied in the realm of trip generation/distribution modelling. Respondents are offered alternative hypothetical destinations and they can indicate which of these destinations are worth a visit. More specifically, the application concerns a scientist who decides what types of colleagues at various locations are worth a visit to carry out collaborative research. The paper proceeds as follows. Section 2 states the problem in formal terms and introduces the estimation method. In Section 3 the data of our application are presented. In Section 4 the results of the estimations are given, and in Section 5 some conclusions are drawn.