Behavioral Modeling of Urban Freight Transport - Sustainable Urban Logistics: Concepts, Methods and Information Systems - page 229

Table 3 Econometric results based on M1

Variable Coefficient St.Err. T-Stat Expected Sign

LUB 0.253 0.048 5.24 +

PLUBF 0.347 0.053 6.51 +

EF -0.699 0.042 -16.44

ASC_Alt1 0.824 0.154 5.32 *

ASC_Alt2 0.657 0.136 4.82 *

in the probability of finding them free has a positive impact on retailers' utility. On
the contrary, an increase in EF has a negative impact on retailers' utility. M1 also
includes two ASCs for the unlabeled hypothetical cases (ASC_Alt1, ASC_Alt2)
whose coefficients represent the overall alternative impact on retailers' utility
when all the coefficients of the other attributes have a zero value. In our case,
results show that, there is an a priori evaluation against the SQ (ASC_Alt3 has a
negative sign) and, after conducting a Wald test for ASC_Alt1 and ASC_Alt2, we
cannot reject the null that the difference between the two coefficients is different
from zero. In summary, one can affirm that ASC_Alt1 and ASC_Alt2 have a
positive, but undistinguished between them, effect on utility. Furthermore, it is also
interesting to note that the ASC inclusion in the model not only substantially
increased the model fit but also provided more realistic interpretation of the
parameters.
The normalization adopted for the explanatory variables allows us to compare
the estimated coefficients of the attributes considered. One can notice that tariff
plays the lion part in explaining retailers' preferences. In fact the EF's coefficient
is more than double the sum of LUB and PUBF coefficients. This result is further
reinforced by looking at the t-stat of each of the variables considered that testify
EF's coefficient is, almost for sure (t-stat 16.44), different from zero even if LUB
and PLUBF coefficients are highly significant too (respectively t-stat 5.24 and
6.51).
M2, reported in Table 4 , differs from M1 in the treatment of the variables
which, in this case, are effects coded. 10 The different coding aims at detecting
possible non-linearities in the explanatory variables' effects. In fact, the estimation
of a single parameter for a given attribute will give rise to a linear estimate (i.e.
slope) and we generically refer to these estimates as linear estimates (M1). An
attribute's impact can be estimated with two dummy (or effects) parameters, which
are usually referred to as a quadratic estimate or higher degree dummy (or effects)
parameters which are also referred to as polynomial of degree L-1 estimates (with
L denoting the number of dummy or effects parameters). In more detail, one can
affirm that the more complex the part-worth utility function, the more advisable is
to move to more articulated coding structures capable of recovering the necessary
data to estimate the more complex non-linear relationships.

10 For a clear description of effects coding the explanatory variable please refer to Hensher et al.
( 2005 ), pp 119-121.