reliability of at least 90 percent are described. The parameters shown in the left part (i.e. MNL
model) are used as basis for mixed logit (ML) estimation. Mixed logit models are examples of
discrete choice models that can test for the possibility that pairs of alternatives in the choice
set are correlated to varying degrees. For example, a bus and train may have a common
unobserved attribute (e.g. comfort), which makes them more similar (i.e. more correlated)
than either is to the car. These choice models can also allow for differences in variances of the
unobserved effects (Louvière et al., 2000). The ML model does not suffer from the IIA (i.e.
independence from irrelevant alternatives) en IID (i.e. independently and identically
distributed) restrictions with which the MNL model is confronted (Louvière et al., 2000;
Train, 2003). The model is therefore seen as a better and more advanced estimation model
than the MNL model (see also Louvière et al., 2000; Train, 2003). Results based on ML
estimation are also presented in table 11. The same parameters were taken into account as in
the MNL-case. Additionally, the variables ‘number of bedrooms’, ‘monthly cost’, ‘big city
and small town’ and the ‘travel time and cost’ variables were tested on randomness. For each
of these variables, triangular and normal distributions (representing amongst other things
unobserved heterogeneity in preferences) were applied. The best fitting model, looking at
significance of coefficients, is presented in table 11, in which a triangular distribution was
used for the monthly housing cost, travel cost and travel time coefficients and a normal
distribution was applied for the ‘big city’ variable. Because of the superiority of the mixed
logit estimation procedure, the description of results in this section is mainly based on the
mixed logit outcomes.
Each respondent within the experiment made 9 choices. The presence of multiple
observations (i.e. 9 choices per individual) on stated choice responses for each sampled
individual means that a potential for correlated responses across observations exist. This is a
violation of the independence of observations assumption in the classical choice model
estimation (Hensher and Greene, 2003). The possibly existing correlation can be the product
of many sources including the commonality of socio-economic descriptors that do not vary
across the choice situations for a given sampled individual and the sequencing of offered
choice situations that results in mixtures of learning and inertia effects, amongst other
possible influences on choice response. Through the applied estimation procedure, these
possibly existing correlation effects were accounted for.
19