cance of the parameter distributions.7 By contrast, the partitioning of the choice
set using the error components appears to be an essential model specification.
Not only is the standard deviation on “annoying” alternatives highly significant,
indicating that the utilities of the respective retrofit alternatives are correlated,
but there is also clear-cut evidence of a significant improvement in the fit of the
model compared to the models that omit the error components.8 We thus con-
clude that the error components logit model is the superior choice for these data,
and we proceed by calculating the respondent’s marginal willingness-to-pay for
energy savings using the coefficient estimates from this model.
3.2 Marginal willingness-to-pay and its Distribution
The household’s WTP for decreasing the building’s primary energy demand by
one kWh can be derived as the marginal rate of substitution between investment
cost and energy savings. For the calculation of the respondent’s marginal WTP
(MWTP), we thus fix the representative utility Vij and take the total derivative
of Equation (8):
dVij = dCij (β1 + βlzil) + dδQij (β2 + βmzim) = 0,
(9)
MWTPi =
dCij _ (β1 + ∑ l βlzim )
—----- —--∙-------------.
dδQij (β2 + ∑m βmzil)
Hence, individual MWTP can be expressed as the ratio of the cost and energy-
saving coefficients, including their interaction effects.
7A likelihood ratio (LR) test of the conditional logit model without error components against
the random parameter logit without error components and two degrees of freedom yields a LR
statistic of 0 (p=0.5). The corresponding LR statistic from the models with error components
is 2 (p=0.184).
8A likelihood ratio (LR) test of the conditional logit model without random parameters
against the error components logit without random parameters and two degrees of freedom
yields a LR statistic of 38 (p less than 0.0001). The corresponding LR statistic from the models
with random parameters is 40 (p less than 0.001).
17