2005 A 1 Eng Valuing Noise Level Reductions in A Residential Location Context 18 P
2005 A 1 Eng Valuing Noise Level Reductions in A Residential Location Context 18 P
www.elsevier.com/locate/trd
Abstract
Although noise levels in urban areas frequently surpass the designated norms the consequences on health
are only beginning to be examined. A stated preference experiment is used to estimate the willingnessto-pay for reducing noise levels in a group-based residential location context. The experiment considers
variations in travel time to work, monthly house rent, sun orientation of the dwelling and subjective noise
level inside it; objective noise levels are also measured after the experiment. Multinomial and mixed logit
models are estimated based on a consistent microeconomic framework, including non-linear utility functions and allowing for various stratications of the data. The more exible models allow for the treatment
of repeated observations problem common to stated preference data and provide a better t to the data,
although willingness-to-pay results remain almost invariant. Subjective values of time derived from these
models are consistent with previous values obtained in the country, giving support to the experimental
design and quality of the sample.
2005 Elsevier Ltd. All rights reserved.
Keywords: Noise valuation; Willingness-to-pay; Stated preference; Discrete choice modelling; Value of time
1. Introduction
As urban transport systems grow externalities often become important. The social evaluation of
transport projects usually includes the benets associated with time savings, however the benets
*
Corresponding author. Tel.: +56 2 686 4822; fax: +56 2 553 0281.
E-mail address: [email protected] (J. de D. Ortuzar).
1361-9209/$ - see front matter 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trd.2005.04.004
306
of reducing the number and severity of accidents are only considered seriously in rst world nations, and the potentially large benets associated with reductions in pollution and noise are seldom considered even in the developed world.
There are at least two ways to incorporate the social valuation of external eects. The rst attempts to quantify the change in the aggregate product value in terms of the impact on social or
individual productivitythe social accounting or shadow price approach. The second involves
estimating the perception of users about the damage inicted, either through observations of their
actions (revealed preference) or through analysis of their stated willingness-to-pay (WTP), the socalled questionnaire methods.1
Here we deal with the problem of valuing reductions in noise levels. Noise is an endemic problem in large metropolis and has received relatively little attention as a potential health hazard in
the developing world. For example, measurements carried out by the Environmental Health Service in Chile (SESMA, 1999) have shown that in most areas of its capital, Santiago, the relatively
lax Chilean norm of a maximum of 55 dB(A) in residential areas, is violated during many hours of
the day.
This is, however, a complex subject due to the diused nature of the contribution from each
agent and the complex causal relation between noise and health. Several methods have been applied to value noise eects in the past, including hedonic pricing (Vainio, 2001) and contingent valuation (Feitelson et al., 1996; Bjorner, 2004)a comprehensive review is provided by Navrud
(2002). Nonetheless, these methods have well-known deciencies (Hausman, 1993; Azqueta, 1994).
On the other hand, stated preference (SP) methods have been widely used in marketing and in
transport research to identify responses to choice situations that are not clearly revealed in the
market. As such, they appear to be a promising tool to estimate monetary values for improvements to the environment. Here we report the rst experience in using SP data to estimate the
WTP for reducing noise levels in Latin America. Few studies of this type have been reported before (Saelensminde, 1999a; Daniels and Hensher, 2000; Wardman and Bristow, 2004).
Household location choice is studied for theoretical reasons. In valuation experiments, it is necessary to induce a direct perception of damages by the individual and this cannot be achieved in a
mode or route choice context. To interpret the results of the SP approach, this work is based on
the microeconomic residential location model developed by Perez et al. (2003). The group-based
approach (i.e. making the whole family participants of the SP game) is generally seen as superior
to more traditional individualbased studies (Molin et al., 1999) and has been used to measure
WTP for improved accessibility and for lower environmental pollution (Ortuzar et al., 2000; Ortuzar and Rodriguez, 2002).
2. Survey design
As valuation of noise reductions is not an everyday consideration for individuals it is important
to be careful in selecting an appropriate choice context for doing this, and the specic way, the n
variable metric, in which it would be presented in the survey.
1
Freeman (1993) and more recently Arsenio (2002), examine and classify the various methods that have been
proposed in the literature.
307
308
louder than others). For this, participants were asked to rank using a ve-point scale. The results,
however, did not show a clear pattern and none of the respondents came close to the objective data
in their assessments.
Laboratory experiences were discarded, although they would allow the simulation of a wide
range of situations, because the experiments are costly and there is no way to know if respondents
feel at home in this simulation or if they would be really annoyed by that noise level in practice
(Arsenio et al., 2000). Eventually, a traditional 10-point rating scale was adopted. Finally, in the
case of sun orientation cardinal points are used; this is not only the natural way of presenting the
attribute but it was easily understood by participants and conrmed at the subsequent pre-tests.
2.3. Experimental design and statistical design
After selecting the attributes and their representation, the next step is to select their numbers of
levels. Although more levels allow testing for non-linearities, the number of choice situations increases and so does respondent burden (Ampt, 2003; Caussade et al., 2005).
The factorial design chosen for this survey was a 24 experiment (Street et al., 2001). A full factorial needs 16 choice situations but has been found in past work to test respondents patience.
Two blocks with eight treatments each are thus used, confounding the four-way interaction
(A B C D). Although two-way and three-way interactions can be estimated, quadratic eects
cannot be estimated because this represents a 2k design (Louviere et al., 2000).
Two blocks of eight treatments are generated by solving3:
Block 1: A + B + C + D = 0 (mod 2)
Block 2: A + B + C + D = 1 (mod 2)
0 0 0 0 (alternative 1)
0 0 0 1 (alternative 6)
1 1 0 0 (alternative 8)
0 0 1 0 (alternative 5)
1 0 0 1 (alternative 7)
0 1 0 0 (alternative 2)
1 0 1 0 (alternative 3)
1 0 0 0 (alternative 3)
0 1 1 0 (alternative 5)
1 1 1 0 (alternative 4)
0 1 0 1 (alternative 4)
1 1 0 1 (alternative 7)
0 0 1 1 (alternative 2)
1 0 1 1 (alternative 8)
1 1 1 1 (alternative 6)
0 1 1 1 (alternative 1)a
a
This alternative was modied, as seen in Table 1, and replaced by (0 0 0 0) to introduce more
realism at the cost of loosing complete orthogonality.
Table 1 details these results. If a respondent answers block 1 seriously, the person should rank
alternative 8 rst and alternative 2 last. On the other hand, a respondent answering block 2 seriously should put alternative 3 before alternatives 1, 5, 6 and 8. These are just examples; blocks 1
and 2 imply 12 and nine dominated pairs of alternatives.4 This enables checking of the data for
inconsistencies (Saelensminde, 2001, 2002).
3
Variables with two levels are represented by zero at the lower level and by one at the higher level.
Dominance appears when every attribute of an alternative is in an equal or better level than those of the other;
therefore, the former should be placed always higher than the latter in the ranking.
4
309
Table 1
Attributes levels in each block
Alternative
1
2
3
4
5
6
7
8
Block 1
Block 2
Time
Noise
Sun
Rent
Time
Noise
Sun
Rent
High
High
Low
High
High
Low
Low
Low
High
High
High
Low
Low
Low
High
Low
Best
Worst
Worst
Best
Worst
Worst
Best
Best
Low
High
Low
High
Low
High
High
Low
High
High
Low
Low
High
High
Low
Low
High
Low
High
Low
High
High
Low
High
Best
Best
Best
Worst
Worst
Best
Best
Worst
Low
Low
Low
Low
Low
High
High
High
Each interviewer was provided with a survey manual containing the precise set of questions to be formulated and
exact denitions of the data required; e.g., the manual specied that an adult family member should be the rst
contacted to ask for the general household information.
310
Table 2
Variation levels for each attribute
Attributes
Time travel to work (TTW)
Monthly rent (MR)
Noise level (NL)
Level 1 (%)
15
10
15
15
10
15
rst stage involved obtaining details of the main characteristics of the dwelling and basic information about family members. Socio-economic characteristics required were: rst name, relation to
household head, gender, age, educational level, possession of a driving license, and occupation.
The household data were: borough where the dwelling was located, nearest street intersection,
monthly rent/mortgage paid, origin of this money, number of household vehicles and family income. The interviewer also gathered trip data from every worker in the family: modes used, borough where the work place was located, nearest street intersection to the work-place, current
travel time and weekly number of trips to work. Finally, the family was asked about which sun
orientations were considered best and worst, and requested them to grade the current level of
noise inside the dwelling according to the 10-point scale.
The second stage involved a second visit two days later where a customized SP exercise was presented. The experiment was generated on the basis of the data collected at the rst stage and included a set of complementary questions. The interviewer made an introductory description of the
SP context and varying attributes and then delivered 8 cards representing dierent residential
locations (Fig. 1). The family was asked to rank these cards and after completing the process
the interviewer asked questions designed to see if the family members had played the game consistently, how important they considered the variable noise, and if the attribute levels had been
considered realistic (Galilea, 2002).
Because the main purpose of the exercise was to value reductions in noise it was felt important
to measure the noise level of each dwelling. Due to the high cost of this it was decided to interview
families living only in predetermined buildings. An important step was to get permission to do the
survey in each of these buildings; a registry of noise levels was oered in return.
Finally 150 ats were surveyed in nine buildings in dierent areas of Santiago (four in the high
income district, two in a medium-high income sector and three on the boundary between middle
and low income sectors). The buildings were selected on the basis of their socio-economic characteristics and noise levels. The number of households interviewed at each building varied between 3 and 4 in high-income buildings to between 27 and 33 in low-medium income buildings.6
The socio-economic information of the sample is representative of middle and high-income
people in Santiago. Table 3 shows the distribution by rent paid and family income; 63.5% of
households answering the income question stated an income level over 850,000 Ch$/month. This
The sampling strategy was to get as many households who have moved in the last year as possible in each building.
No record of detailed response rates was kept, although of 12 preliminary selected buildings, we were given permission
to conduct our survey in 9.
311
Table 3
Rent/mortgage distribution by family income
Rent/mortgage
(103 Ch$/month)
099
100199
200299
300399
Over 400
Total
Total
351550
551850
8511250
Over 1250
No answer
3
5
2
2
15
2
1
1
24
4
1
1
21
9
4
1
11
14
11
5
1
4
3
4
1
9
80
34
21
6
10
20
30
35
42
13
150
is among the highest 10% of income in the countrythe minimum wage being a little over
100,000 Ch$/month (at the time $1 = Ch$650).
The work on measuring noise levels started as soon as the second visit was over for all households. Attempts were made to measure inside ats to spot dierences resulting from street-level
height or orientation, however, only 64% of ats were measuredthe rest either refused entrance
or nobody was in when the noise specialist arrived.
3.2. Consistency and lexicographical behaviour
There is a need to detect observations that are not internally consistent or that do not correspond to the assumed population behaviour. Those households that did play the SP game as carefully as required are rst identied. A second task was to identify those cases where responses
suggested that the household decision strategy could be inconsistent with the compensatory decision making protocol assumed.
312
Comparing eight alternatives with four attributes each is not easy and it is to be expected that
some households would make mistakes and be inconsistent. A maximum of two inconsistent responses per household is seen a as a reasonable indicator of a carefully completed exercise. This
led to the elimination of 65 individual responses and all observations from 18 households with
three or more inconsistencies.
Forty households exhibiting potentially lexicographical behaviour (i.e. ranking on the basis of a
single attribute) were detected; 16 on the attribute rent, 23 on sun orientation, and 1 on travel time
to work.7 This is a slightly lower proportion than reported in previous studies (Saelensminde,
1999b; Ortuzar and Rodriguez, 2002; Rizzi and Ortuzar, 2003; Iraguen and Ortuzar, 2004).
The relatively high presence of lexicographical households with respect to sun orientation may
be due either to a real family concern for this attribute or, and more probably, to the use of
the best and worst orientations as levels for this attribute which oered only extremes.
The potentially lexicographical responses were kept in the nal estimation process for various
reasons. First, there is never a certainty that respondents are truly lexicographic (see the discussion by Hojman et al., 2004). Second, their inclusion allows to compare SP results with those
of revealed preference (RP) surveys, where lexicographical answers cannot be detected. Finally,
in this case the models including lexicographical answers t the data as well as those excluding
them.
where TTWhi is the travel time to work by individual h from location i (minutes per trip) and fih is
the frequency of trips to work by individual h, from location i (trips per week).
We label these potentially, because it could be that although they actually choose in a compensatory fashion, the
levels of the experiment do not imply a real trade-o for them.
8
Ortuzar and Willumsen (2001) provide summary of the economic and statistical background for analysing discrete
choice data.
313
Two MNL models were estimated; the rst included all consistent responses by every household,
and the second excluded households exhibiting potential lexicographical behaviour. The maximum likelihood results show that both models have a satisfactory adjustment in comparison to
the market shares model, correct signs and signicant parameters (Galilea, 2002). This diers
from previous experience when models excluding potentially lexicographical respondents had
been found superior (Rizzi and Ortuzar, 2003; Iraguen and Ortuzar, 2004). For this reason models
for the complete sample are considered.
Segmentation is tested to quantify the inuence of certain household characteristics in the valuation of attributes, but only at ownership was signicant. To identify the best specication at
each stratum, a step-by-step methodology proposed by Ortuzar and Rodriguez (2002) is used:
1. Estimate a general model with dierent parameters in both segments; identify statistical similarity between parameters considering magnitudes and t-tests.
2. Estimate a restricted model under the null hypothesis that the most similar pair of parameters can be replaced by a single parameter in the estimation process.
3. Perform a likelihood-ratio test to ensure that both models (general and restricted) are statistically equivalent and that, consequently, there is a gain in parsimony.
4. Recheck parameter similarity, identify the closest pair of specic parameters remaining and
go back to step 3.
5. Stop when no potential pair similarities or model reductions are found.
Owners and tenants are considered. Although their numbers were not equivalent (43 and 107,
respectively), the results in Table 4 are solid. A likelihood ratio-test shows that the stratied structure MNL-2 is signicantly superior to MNL-1, the original model without stratication.
WTP values and their 95% condence intervals are calculated following Armstrong et al. (2000)
(Table 5). The subjective values of time (SVT) are in complete agreement with values estimated in
previous studies (Ortuzar et al., 2000; Perez et al., 2003) and this gives credibility to the experiment, i.e. not only is it well designed and understood, but also respondents answered with seriously. Secondly, the subjective value of reducing noise (SVN) has the same order of magnitude
as the money needed to double-glaze in dwellings; these are values based on peoples perceptions
of noise and not objective values.
On the other hand, model MNL-2 implies that at owners value sun orientation and travel time
higher than tenants, but their perception of noise and rent values are not signicantly dierent;
consistent with the ndings of Bjorner (2004). The rst result can be explained by considering that
when a person is buying an apartment there is a longer-term commitment than when choosing one
to rent, so sun orientation in a warm country should be more important. But the much larger
parameter associated to owners travel times to work is probably due to their higher income level.
Finally, although the condence intervals for the SVT overlap, the interval for the tenants does
not include the owners point estimate. Their condence interval being wider could be explained
314
Table 4
Model stratied by dwellings ownership
Attributes
Parameters (t-test)
MNL-1
MNL-2
RM (10 Ch$/month)
0.043 (16.3)
0.0436 (16.4)
NL
0.660 (14.3)
0.665 (14.3)
Tenants
SUN
1.566 (13.5)
1.697 (16.2)
Owners
2.142 (10.4)
0.006 (6.1)
Tenants
0.006 (7.3)
TTW (min/week)
0.009 (4.5)
Owners
967.03
0.213
859
l(h)
q2c
Sample size
963.28
0.216
859
Table 5
Subjective values by dwellings ownership
Attributes
MNL-1
MNL-2
23.54 (20.2627.24)
23.45 (20.1727.09)
Tenants
TTW (US$/h)
2.81 (1.933.70)
3.10 (2.303.92)
Owners
4.43 (2.536.37)
by the fact that, although most owners had a higher income level than the tenants, this was not
true for every owner.
4.2. MNL models including interaction eects9
The factorial design allows estimation of models involving interactions between variables. The
specication searches began with a general model including all main eects, two-way and threeway interactions and then less signicant eects were taken out, one by one, until a model was
reached with signicant variables and intuitively correct signs. The results were checked by choosing dierent ways to take out the less signicant eects; this model had a signicantly better t
than MNL-1 (Galilea, 2002).
9
All models with interactions and with parameterised main eects were estimated deviating each variable by its mean.
This helps estimating the condence intervals and does not change the results. The variable means were calculated based
on the attributes eectively available for each observation at the estimation process.
315
On the other hand, Eq. (2) does not allow incorporation of individual tastes in a MNL. One
way of doing this is to parameterise the coecients of the main eect variables by means of
the individuals socio-economic and journey characteristics:
!
!
!
X
X
X
al slj RMij b0
bl slj NLij c0
kl slj SUNij
V ij a0
l
d0
!
dl slj TTWij .
In contrast with the traditional specication of socio-economic variables, (3) applies to both alternatives; also, since the same additional variable can be related to more than one attribute, it can
be specied with dierent coecients in each case. Therefore, as every individual has dierent socio-economic and journey characteristics, each may end up with dierent valuations for the same
attributes. Rizzi and Ortuzar (2003) provide a microeconomic rationale for Eq. (3).
The binary variable slj represents socio-economic (SE) feature l of individual j. This is an interesting way of incorporating additional variables and allow us to use additional individual data to
estimate WTP. As there may be dierent coecients for each attribute depending on the special
characteristics of each household, this specication allows estimating models that are almost
unique to each household helping to reduce the problem of taste variations. After a specication
search the most signicant variables were (Galilea, 2002):
No. people/IncomeRM: Number of household members divided by family income level; as it was
added to the rent coecient its value should be negative, if the number of members increases or
if income decreases (ceteris paribus), an increase in rent should aect them more.
Floor2NL : Takes the quadratic value of the oor where the apartment is located. Its value should
be negative, because the noise level is higher in the ground oor (closer to the source of noise)
and in the upper oors (no shield from other houses or smaller buildings), so people living in
these oors should be more sensitive to higher noise levels.
OwnerSUN: Takes the value of one if the household owns the at. Its value should be positive
because owners attach more importance to the sun orientation than the rest.
The best model (MNL-4) incorporating both interactions and parameterised main eects is
shown in Table 6. To compare it with the linear MNL we had to estimate a new model
(MNL-3) using only data from families reporting their income and who passed the consistency
test (119 households). As can be seen, all parameters in both models have correct signs but a likelihood-ratio test rejects comfortably the null hypothesis that both models are equivalent (i.e.
LR 33.7 > v26;95% 12.6).
The results indicate that, for example, a household with a given income level will increase its
valuation of the rent in 0.031 (i.e. almost 85%) for each new member added to it. On the other
hand, at owners value sun orientation 36% higher (i.e. 1.674 plus 0.609 over 1.674) than do
tenants.
It is surprising to nd more three-way than two-way eects present in the preferred specication, as it has been generally assumed that the latter explain a greater proportion of the data
316
Table 6
Model with interactions and parameterised main eects
Attributes
Parameters (t-test)
MNL-3
MNL-4
RM (10 Ch$)
No people/IncomeRM
NL
Floor2NL
SUN
OwnerSUN
TTW (min)
RM NL
RM SUN TTW
NL SUN TTW
0.0447 (15.9)
0.6789 (13.8)
1.743 (15.7)
0.00621 (6.9)
0.0360 (6.8)
0.0307 (2.6)
0.6470 (10.3)
0.0026 (2.2)
1.6742 (13.5)
0.6086 (2.6)
0.00668 (7.1)
0.00152 (2.7)
0.000025 (2.8)
0.000481 (2.0)
l(h)
q2c
Sample size
877.63
0.222
786
860.78
0.236
786
variation (Louviere et al., 2000); but models with only two-way interactions were consistently
inferior (Galilea, 2002).
Table 7 presents estimates for the subjective values derived from MNL-4. Although the specication is much improved, the results remain almost invariant. These values are computed taking
account of interactions and parameterised eects, so we need to derive utility with respect to the
attributes as
oU
No: people
hRM hNo: people=Income
hRMNL NL
oRM
Income
hRMTTWSUN TTW SUN;
oU
hNL hFloor2 Floor2 hRMNL RM hRMSUNTTW RM SUN
oNL
hNLSUNTTW SUN TTW;
oU
hTTW hNLSUNTTW NL SUN hRMSUNTTW RM SUN.
oTTW
Then, the subjective values of time and noise level are calculated as follows10:
SVT
oU =oTTW 12 000 60
oU=oRM
52
650
SVN
oU=oNL 1000
oU =oRM 650
US$=h;
US$= deg.
7
8
10
In Eqs. (4)(6) we used sample averages for the variables Income, Sun, Floor, Rent (RM) and travel time to work
(TTW).
317
Table 7
Subjective values for models with interactions
Attributes
Noise level (US$/NL per month)
Travel time to work (US$/h)
Subjective value
MNL-3
MNL-4
23.37 (20.1127.04)
2.96 (2.183.74)
23.68 (20.4339.25)
3.00 (2.685.85)
The SVT and SVN values were positive for all individual households, and this is not always the
case in models of this type (Brownstone, 2001; Daniels and Hensher, 2000).
4.3. Estimation of Mixed Logit models
ML models were nally estimated to examine the importance of allowing heterogeneity in individuals tastes explicitly, as well as a correct treatment of the repeated observations problem associated to SP data. In the ML model, apart from the random errors (i.e. white noise) that distribute
independent and identically (iid) Gumbel as in the MNL, the systematic part of the utility function, Viq, may also have randomly distributed parameters. This exible specication allows to
consider heteroscedasticity, correlation and variations in tastes but it is hard to estimate (Train,
2003).
To estimate the ML we used a non-commercial code implemented in GAUSS (it can be downloaded from the web page of Kenneth Train: http://elsa.berkeley.edu/~train). We considered independent Normal distributions for the attributes, as previous experience with multivariate
functions had shown results to vary little but at non-negligible cost (Sillano and Ortuzar,
2005). For the simulated maximum likelihood search we used sequences of 125 Halton numbers
as is usual practice (Train, 2003). A battery of ML models, associated to several of the MNL models estimated during the research is discussed by Galilea (2002); in every case the ML gave a substantially better t to the data than the corresponding MNL. In what follows we just present two
such models, one associated to MNL-1 and another to MNL-4.11 Comparing ML-1 in Table 8
and MNL-1 in Table 4, it can be seen that the mean values of the ML parameters correctly increase in size due to the scale factor eect; as we are allowing for random parameters the white
noise variance (inversely related to the scale factor) decreases signicantly (Sillano and Ortuzar,
2005). Note that this also happens with ML-4 and MNL-4.
The results in Table 8 suggest that ML-4 is superior to MNL-4 (although it is not strictly comparable). Note that although most interactions nally received xed parameters the main eects
remained consistently variable among individuals. The table also shows that ML-4 is closer but
still superior than ML-1.
Table 9 presents the willingness-to-pay point estimates for these nal models. As can be seen,
those for ML-4 are larger than those of ML-1 and MNL-4, but still the condence intervals for
each model contain all the point estimates of the competing functions; nally, again no households were found to have individual subjective values with an incorrect sign. As this is seldom
11
Note that we could not estimate a close enough function as an ML with both interactions and parameterised main
eects did not converge.
318
Table 8
Mixed Logit model including interactions
Attributes
Parameters (t-test)
MNL-4
ML-1
RM
No people/IncomeRM
NL
Floor2NL
SUN
OwnerSUN
TTW
RM NL
RM TTW
NL SUN TTW
RM SUN TTW
0.0360 (6.8)
0.0307 (2.6)
0.6470 (10.3)
0.0026 (2.2)
1.6742 (13.5)
0.6086 (2.6)
0.00668 (7.1)
0.00152 (2.7)
0.000481 (2.0)
0.000025 (2.8)
l(h)
q2c
860.78
0.236
ML-4
Mean
Standard
deviation
Mean
Standard
deviation
0.1398 (5.5)
2.2852 (10.3)
3.718 (9.4)
0.0241 (7.4)
0.0863 (4.7)
1.1504 (8.0)
4.699 (7.9)
0.0137 (5.4)
0.120 (8.6)
2.570 (8.6)
4.566 (7.5)
0.0284 (8.4)
0.00456 (2.6)
0.00007 (3.5)
0.0888 (6.9)
1.283 (7.8)
5.052 (7.0)
0.0097 (5.7)
0.00662 (3.8)
0.00013 (6.5)
787.74
0.359
777.62
0.367
Table 9
Subjective values for Mixed Logit model including interactions
Attributes
Noise level (US$/NL per month)
Travel time to work (US$/h)
Subjective values
MNL-4
ML-1
ML-4
23.68 (20.4339.25)
3.00 (2.685.85)
25.15 (18.5637.63)
3.67 (2.555.63)
33.98 (23.7445.96)
5.18 (3.627.06)
the case, calls are made to include among the model assessment check-list, if the microeconomic
conditions implied by the postulated indirect utility function are violated (Cherchi and Ortuzar,
2003); also the use of special distribution functions to avoid the possibility of incorrect signs
has been advocated (Hensher, 2005; Hess et al., 2005).
4.4. Subjective versus objective (dB) perceptions of noise level
The nal objective is estimation of an objective monetary value for noise levels reductions. To
do this, the 10-point scale subjective values are related to the decibel scale measurements taken at
the 96 dwellings where this is possible. One problem is that the dB(A) measures are generally fairly
high whilst the ranges was not wide (i.e. from 37 to just under 61 dB) implying that many respondents with a low objective noise level report a high grade as their subjective level. So a simple linear regression did not achieve a reasonable t even when separate regressions were estimated for
each building. Only one building, in fact, gave a more satisfactory t but its respondents were too
few for further analyses (Galilea, 2002).
319
To improve the estimation the extra information provided by the households at the interview
stage is incorporated to achieve ceteris paribus conditions. In particular, the results of two questions are used: whether they were aware that their dwelling had a signicant noise level and, if they
thought that the level of noise was an important attribute when searching for a place to live. Thus,
a multiple regression was estimated using the 10-point scale subjective grades as the dependent
variable, and the decibel scale plus two dummies representing Awareness (one, if the household
was aware that the dwelling had a signicant noise level) and Importance (one, if the household
thought that noise level was an important attribute) as independent variables. The results of
the regression are shown in Table 10; they appear quite reasonable. With these results, we are able
to transform the estimated parameter for the noise level, by multiplying it by the coecient for
dB(A).
The transformed noise level parameters allowed us to derive new subjective values of noise level
(SVN) for each model (Table 11). These represent the willingness-to-pay, in US$, associated to
decreasing the noise level inside a dwelling in 1 dB(A) per month. Although the values appear reasonable and consistent, a caveat related to their use in social project evaluation is that there are
other terms and elements that should form part of the total WTP for reducing noise level; for
example, the health costs incurred as an eect on human health because of noise.
The importance of establishing a relationship between noise perception and dB(A) is clear for
cost-benet analysis. For instance, if noise levels in a central business district are high and as a
consequence most households and oces start incorporating double glazing as a personal protection it could be more ecient to coordinate a central double glazing programme to attack the
noise problem, for example, but this should be evaluated. Such a programme could even attract
some other agents which, otherwise, could not aord the cost of double glazing.
Table 10
Multiple regression results for the 10-point scale
Attributes
Parameter
(t-test)
dB(A)
Awareness
Importance
0.0893
2.1295
1.3184
(6.1)
(3.6)
(2.3)
0.5124
96
Table 11
Subjective values for noise level reduction in dB
Model
MNL-1
MNL-2
MNL-3
MNL-4
ML-1
ML-4
Mean
Upper bound
1.81
1.80
1.80
1.82
1.66
2.12
2.10
2.09
2.09
2.12
2.25
3.03
2.43
2.42
2.42
3.51
3.36
4.10
320
As a caveat, we must recall that the relationship established above has to be considered tentative in so far as we did not conduct any kind of external validity of our results. Therefore, it is
debatable whether it could be adopted immediately by the environmental authority. More work
should follow in this direction.
5. Conclusions
The application of SP techniques to a complex problem as valuing noise level reductions in
Chile shows great promise for the use of this methodology in other countries. Two results emerged
from the design stage of the survey: the identication of residential location as an appropriate
experimental framework, and the formulation of a variable metric for noise level (although only
related to family perceptions) that was clearly understood by the participants in the exercise.
We also found that the statistical design was able to represent the respondents preferences for
the variables included in the experiment. This is supported not only by the good general t of the
estimated models, but also by the anticipated parameter signs and reasonable signicance t-tests.
Equally, the subjective values of time obtained from the various models estimated turned out to
be consistent with prior studies. This is, in our opinion, a clear indication that respondents understood the experiment which included two new variables in relation to previous experiences: noise
level and sun orientation. So in spite of its complexity, the SP experiment was able to capture individual preferences adequately.
In terms of results, individual households do not necessarily have linear utility functions. Not
only are several interaction terms signicant but the introduction of additional variables (socioeconomic and related to noise level) aect the coecients of the main-eects variables. These latter variables indicate the signicance of age and income in the valuation of travel time. Further,
the exible and powerful Mixed Logit model easily outperformed the simple MNL; this may be
because it accepts the presence of random taste variations among individuals (which indeed appeared as an eect) or just because it allows consistent treatment of the problem of repeated observations by each individual, which is a feature of stated preference and panel data.
In relation to the estimated values for reducing noise levels, and given the caveats, a conservative value of US$2.12 per decibel per month emerges as corresponding to the lower bound of the
condence interval associated to the best ML model (and also to the point estimate of the various
MNL functions). This value (and most values estimated) appears to be reasonable when compared (although the comparison is per force not strict) with the real costs12 associated to reducing
noise by physical means (i.e. double glazing).
Acknowledgements
The nal version of this paper was written during a six-week stay of the second author at the
Institute for Transport and Logistics Studies, University of Sydney. We wish to thank Ana Mara
12
The cost of double glazing is roughly US$130 per square meter. If, for example, the noise level was near 70 dB(A)
double glazing would reduce it to approximately 40 dB(A). However, there would also be a reduction of approximately
40% in central heating costs.
321
Ivelic for help at the survey design stage, Luis Ignacio Rizzi for his invaluable help throughout the
work and Mauricio Sillano for his pointers in relation to Mixed Logit estimation. Thanks are also
due to Luis Cifuentes, David Hensher, Sergio Jara-Diaz and Huw Williams for their advice. Ken
Button and two anonymous referees were also instrumental in shaping the paper. The research
was supported by the National Fund for Scientic and Technological Development (FONDECYT) through Project 1000616. We are also grateful for the dedicated work of our interviewers
and to those responsible for the noise level measurement campaign.
References
Ampt, E.S., 2003. Respondent burden. In: Stopher, P.R., Jones, P.M. (Eds.), Transport Survey Quality and Innovation.
Pergamon, Amsterdam.
Armstrong, P.M., Garrido, R.A., Ortuzar, J. de D., 2000. Condence intervals to bound the value of time.
Transportation Research E 37, 143161.
Arsenio, E., 2002. The Valuation of Environmental Externalities: A Stated Preference Case Study on Trac Noise in
Lisbon. Ph.D. Thesis, Institute for Transport Studies, University of Leeds.
Arsenio, E., Bristow, A.L., Wardman, M., 2000. An innovative stated preference computer survey model for valuing
noise impacts from road trac. In: Proceedings of Seminar F, European Transport Conference, Cambridge.
Azqueta, D., 1994. Valoracion Economica de la Calidad Ambiental. McGraw-Hill, Madrid (in Spanish).
Bjorner, T.B., 2004. Combining socio-acoustic and contingent valuation surveys to value noise reduction.
Transportation Research D 9, 341356.
Brownstone, D., 2001. Discrete choice modelling for transportation. In: Hensher, D.A. (Ed.), Travel Behaviour
Research: the Leading Edge. Pergamon, Oxford.
Caussade, S., Ortuzar, J. de D., Rizzi, L.I., Hensher, D.A., 2005. Assessing the inuence of design dimensions on stated
choice experiment estimates. Transportation Research B 39, 621640.
Cherchi, E., Ortuzar, J. de D., 2003. Alternative specic variables in non-linear utilities: inuence of correlation,
homoscedasticity and taste variations. In: Proceedings 10th International Conference on Travel Behaviour
Research, Lucerne.
Daniels, R., Hensher, D.A., 2000. Valuation of environmental impacts of transport projects. Journal of Transport
Economics and Policy 34, 189214.
Feitelson, E., Hurd, R., Mudge, R., 1996. The impact of airport noise on willingness to pay for residence.
Transportation Research D 1, 114.
Freeman, A.M., 1993. The Measurement of Environmental and Resource Values: Theory and Methods. Resources for
the Future, Washington, DC.
Galilea, P., 2002. Preferencias Declaradas en la Valoracion del Nivel de Ruido en un Contexto de Eleccion Residencial.
M.Sc. Thesis, Department of Transport Engineering, Ponticia Universidad Catolica de Chile, Santiago (in
Spanish).
Hausman, J.A. (Ed.), 1993. Contingent Valuation: a Critical Assessment. North-Holland, Amsterdam.
Hensher, D.A., 2005. Imposing a globally signed condition on a generalised parameterisation of attributes in choice
models to produce sign compliant willingness to pay distributions (mimeo).
Hess, S., Bierlaire, M., Polak, J.W., 2005. Estimation of value of travel-time savings using mixed logit models.
Transportation Research A 39, 221236.
Hojman, P., Ortuzar, J. de D., Rizzi, L.I., 2004. On the joint valuation of averting fatal victims and serious injuries in
highway accidents. Journal of Safety Research, under review.
Hunt, J.D., McMillan, J.D.P., Abraham, J.E., 1994. Stated preference investigation of inuences on attractiveness of
residential locations. Transportation Research Record 1466, 1735.
Iraguen, P., Ortuzar, J. de D., 2004. Willingness-to-pay for reducing accident risk in urban areas: an internet-based web
page stated preference survey. Accident Analysis and Prevention 36, 513524.
Jones, S., Hensher, D.A., 2003. Predicting nancial distress using reported cash ows: an ordered mixed (random
parameter) logit model. Working Paper, Faculty of Economics and Business, University of Sydney.
322