Behavioral Modeling of Urban Freight Transport - Sustainable Urban Logistics: Concepts, Methods and Information Systems - page 228

vehicles, gasoline, etc.); (6) services (e.g. laundry, flowers, live animals, acces-
sories and animal food, etc.); (7) clothing (cloth, leather, etc.); (8) construction
(e.g. cement, scaffold, chemical products, etc.); (9) other (all those not included in
previous categories). We follow the same classification proposed in Filippi and
Campagna ( 2008 ) for comparison purpose, since it represents still today the most
reliable investigation of urban freight data in Rome.
4 Econometric Results and Policy Implications
This section reports the results of the models 7 estimated for retailers based on the
data obtained via the SRE described in the previous section.
The methodological framework is based on random utility theory, where utility
is modeled as a random variable consisting of both deterministic and stochastic
part. The former is a function, linear in its parameters, of the fundamental attri-
butes, while the latter is the random term. Different assumptions about the dis-
tribution of the random term imply different discrete choice models that can be
used to analyze the gathered choice data with the purpose of estimating the
parameters associated with the attributes. In the early 1970s Mc Fadden ( 1974 )
developed MNL which is derived from the assumption that the error terms of the
utility functions are independent and identically Gumbel distributed. MNL has
many interesting and much appreciated advantages (closed form, ease of inter-
pretation, etc.) is also characterized by relevant drawbacks linked to preference
homogeneity assumption across respondents. Even if confounded for the scale, the
estimated parameter represents the marginal utility of each attribute variation and
implies an equal taste for all agents for the given attribute. 8
The first model (M1), employing a MNL specification, 9 utilizes all attributes as
linear and normalized while the second (M2) adopts an effects coding for the
variables in order to investigate potential non-linear effects of the different levels
of the explanatory variables.
M1, reported in Table 3 , employing just normalized variables, provides inter-
esting results and also shows a good fit of the model (adj. Rho 2 = 0.142; 5 Coeff.).
All the coefficients are statistically significant and with the expected sign with the
exception of the two alternative specific constants (ASCs) for which there was no
strong a priori concerning the sign. In particular LUB and PLUBF have a positive
coefficient since an increase in either the number of loading and unloading bays or
7 The models are estimated using NLOGIT 4.0.
8 For a detailed discussion of the methodological framework and possible applications of
discrete choice models see, for example, Marcucci ( 2005 ); Gatta ( 2006 ).
9 We just recall that a MNL specification of the model implies an implicit assumption
concerning the independence from irrelevant alternatives. In other words, it is assumed the
unobserved effects homogeneously impact all the alternatives in the same way that is equivalent
to hypothesizing that the error component is identically and independently distributed.