ANALYSIS OF THE STATED PREFERENCE-CHOICE DATA
(7.20) Several dummy incremental effects were specified such as in the example above:
7.7. COMPARISON OF MODELS 1 Goodness-of-fit Measures
7.8.5 Confidence Intervals for the M arginal Values of Quiet
The derivation o f confidence intervals o f the marginal values o f quiet (point estimates) is an ongoing research area. Considering the environmental and stated preference literature reviewed, appropriate formulae for setting confidence intervals was already derived to bound the value o f time (Armstrong et al. 2001). In this study, the quiet and cost variable interacts with other segmenting variables (additional variables), and for this reason the proposed
methods by Armstrong et al. (2001) cannot be applied. As suggested (Ettema et al. 1997;
Armstrong et al. 2001) the confidence intervals can be derived alternatively by using multivariate normal simulation.
The m ethodology followed in the present research to set the confidence intervals lor the marginal values estimated was conducted using simulation o f multivariate normal variates, i.e. the param eters o f the best fit model estimated (mixed logit) were computed from a intervals for the marginal values o f quiet presented earliei (Tables 7.17 and 7.1S) considering the model with higher fit with the data estimated.
Table 7.19: Confidence Intervals for a Situation of Im provem ent (Gain) in Quiet
Table 7.20: Confidence Intervals for a Situation of Improvem ent (Gain) in Quiet
Unit values in 1999 Escudos (1 Euro— 200,482 Escudos).
Table 7.21: Confidence Intervals for a Situation of Deterioration (Loss) in Quiet (Flat Exposure Fronting the Main Road).
Unit values are in 1999 Escudos (1 Euro— 200,482 Escudos).
Table 7.22: Confidence Intervals for a Situation o f Deterioration (Loss) in Quiet
Unit values in 1999 Escudos ( Euro= 200,482 Escudos).
7.9 C O N C L U SIO N S
The modelling work conducted is novel in considering householders’ heterogeneity (nature and extent) o f preferences on the marginal valuations o f traffic noise externalities in the home. A range o f variables (situational, socio-economic, behavioural, attitudinal) was collected by means o f the SP CAPI surveys. Because the SP-choice context used in this research is also novel, the range o f community noise studies reviewed were not conclusive with respect to the expected impacts o f most variables. Therefore, the econometric research allowed a certain degree o f flexibility by testing several alternative functional forms for the explanatory variables in each case.
The stated preference-choice data was driven by respondents’ perceptions o f the internal noise levels indoors and other qualitative attributes intrinsic to apartments and blocks they were familiar with. A range o f other variables collected during the main survey (noise levels inside apartments and outside; socio-economic variables related to each respondent, etc) served to build a range o f multinomial logit models with additional variables o f the main influential variables on quiet (noise). Three types o f MNL-INT models were developed following a common step-by-step methodology and econometric principles, considering the quiet (noise) variables expressed as perceived levels (as rated), as the equivalent Leq dB(A) measures taken indoors and as the equivalent Leq dB(A) measures outdoors, taken Irom each apartment window. The model that performed best was identified, considering the respective
goodness-of-fit m easures, ability to capture the main influential variables (interaction effects with quiet/noise) and plausibility o f marginal values o f quiet (noise).
Findings showed that the model based on perceptions was statistically superior to both the models based on physical noise measures. Overall, the models based on perceptions and the physical noise m easures indoors have captured the most influential effects on the marginal values o f quiet (noise). This finding pointed out the importance o f non-acoustical factors besides the physical noise measures in explaining preferences for quiet. Using realistic noise changes’ situations, the marginal values o f noise per unit o f rating converged to those as m easured, in the sense that losses were on average much higher valued than gains in the same apartment situations, and almost equally valued as gains in similar situations involving a noise change along the same facade.
The implication o f this finding for future noise valuation studies in a similar SP choice context is that w henever data on respondents’ perceptions do not exist, then the physical noise measures indoors have necessarily to be used to get plausible marginal values of noise.
If the physical noise measures indoors cannot be taken (or are too costly), then the noise m easures indoors need to be computed by mixed engineering and acoustics approaches, e.g.
taking the predicted noise levels outdoors (in each exposed floor) and correcting those for the planned insulation conditions (fa9ade characteristic such as materials and window types, area o f windows per fafade, etc).
The marginal value function is asymmetric for the models based on perceptions o f quiet and the equivalent Leq dB(A) measures indoors. Marginal improvements in quiet (noise) are less valued than deteriorations. This finding is in line with recent studies o f marketing science and psychology. From these studies it was expected that losses have a greater impact on individuals’ utility than gains, but for other types o f goods. To my knowledge this is the first study where an asymmetric marginal value function is tested for the case o f noise in the residential context, confirming in a large extent the reference-dependence theory.
The finding that the value o f quiet function is asymmetric has a direct implication in terms of transport planning and environmental impact assessment: if two transport projects (e.g.
construction o f alternative road versus public transport) are supposed to have the same absolute impact in terms o f the noise levels (e.g a 10 dB(A) deterioration in the noise levels from the status quo and a 10 dB(A) improvement in the noise levels, respectively), it can be
said that if the former project is chosen it will produce a much greater change in utility to the exposed householders in comparison to the other option.
The standard multinomial logit models with additional variables have represented the observed heterogeneity on the marginal values o f quiet (noise), considering the main influential variables tested. The inclusion o f these additional variables significantly increased the explanatory pow er o f the base model in all cases. The key explanatory variables were the general flat exposure to main road, base level o f noise experienced, size o f noise changes, adjusted household income per person and floor number, number o f years living at the apartment, gender and base monthly payment as housing service charge. The income elasticity o f marginal values o f quiet was found to be less than one (0.5). Considering realistic mean noise changes, a one unit o f perceived loss (gain) in quiet was marginally valued in the range 671 (403) to 1145 (1052) Escudos per month per household (1999 prices). One dB(A) increase (decrease) was valued between 277 (197) to 451 (370) Escudos per month per household. One unit o f perceived gain and loss (as rated) was tound equivalent on average to 2.9 and 2.7 dB(A) respectively.
The MNL-INT specification was compared with a M L specification. In the ML models tested, random (unobserved) heterogeneity over the deterministic (observed) heterogeneity was allowed. The issue o f taste variation across the sampled individuals could be understood in a more comprehensive way, following a step-by-step methodology for finding the best distribution for each random coefficient.
It was found that the mixed logit specifications provided the best fit with the data. This allows the curvature o f the indirect utility function to vary across individuals o f the same observed heterogeneity (additional variables ol the influential variables ol the standard M NL-INT). The omission o f random parameters (standard MNL-INT) was shown to conduct to higher value o f quiet estimates in some situations and to lower values in others.
Overall, the bias could be considered o f small magnitude. Considering estimated distributions o f the random parameters, it was shown that the respective standard deviations were highly significant. This indicated that there exists a significant heterogeneity at the individual level across households with the same observed influential characteristics.