Accommodating attribute processing strategies in stated choice

Accommodating attribute processing
strategies in stated choice analysis: do
respondents do what they say they do?
Danny Campbella
a
Victoria S. Lorimerb
Institute for a Sustainable World, Queen’s University Belfast, Email: [email protected]
b
Gibson Institute, Queen’s University Belfast, Email: [email protected]
European Association of Environmental and Resource Economists Annual
Conference, Amsterdam, June 2009
Abstract
Data from a discrete choice experiment on restoration of environmental damage
caused by illegal dumping are used to investigate the implications of attribute
processing strategies. Using random parameters logit models, where we account
for heterogeneity in both preferences and attribute processing strategies, this
paper explores whether the responses to follow-up questions on attribute nonattendance are consistent with those picked up analytically. Results from the
analysis provide evidence that allowing for rationally adaptive behaviour leads
to significant improvements in goodness-of-fit and has repercussions for willingness to pay estimation and policy appraisal. Our findings also call into question
the accuracy of respondent’s statements of attribute non-attendance.
Keywords: attribute processing strategies, discrete choice experiments, random
parameters logit model
JEL classifications: C25, Q24, Q51, Q53
1
Introduction
Since its introduction by Louviere and Hensher (1982) and Louviere and Woodworth (1983) there has been a growing number of studies using the discrete
choice experiment methodology. Discrete choice experiments are appealing as
value derivation techniques because they are consistent with the Lancasterian
microeconomic approach (Lancaster, 1966), whereby individuals derive utility
from the different characteristics, or attributes, that a good possesses, rather than
directly from the good per se. Accordingly, a change in the level of an attribute
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
describing a given alternative may cause the respondent to favour that alternative
over another that is perceived as providing an inferior combination of attributes.
In discrete choice experiments, respondents are asked to select their preferred
alternative from a given set (the choice set), and are typically asked to perform a
sequence of such choices (Alpı́zar et al., 2003) giving rise to a panel of discrete
choices. Experimental design theory is used to construct the alternatives, which
are defined in terms of their attributes and the levels these attributes could take
(see Louviere et al., 2000). This type of analysis has been widely used to derive
welfare estimates for ecological and environmental goods.
Despite the widespread use of discrete choice experiments, evidence presented in numerous papers indicates that responses made by some respondents
are inconsistent with standard economic assumptions. This has led to continuing
scepticism about stated preference approaches. As noted by Sugden (2005), if
stated preference methods are to become more generally accepted, a defensible
strategy for coping with such anomalies is essential. This paper is intended to
contribute to this active debate.
A basic assumption, which gives rise to the continuity axiom, within the
discrete choice experiment framework is that of unlimited substitutability between the attributes used to succinctly describe the alternatives in the choice
set. This implies passive bounded rationality, whereby respondents make tradeoffs between all attributes across each of the alternatives, and are expected to
choose their most preferred alternative. Thus, the continuity axiom rules out
rationally adaptive behaviour, whereby respondents focus solely on a subset of
attributes, ignoring all other differences between the alternatives. Ignoring attributes in the choice set implies non-compensatory behaviour because no matter
how much an attribute level is improved—if the attribute itself is ignored by the
respondent—then such improvement will fail to compensate for worsening in the
levels of other attributes (e.g., Spash, 2000; Rekola, 2003; Sælensminde, 2002;
Lockwood, 1996). Therefore, respondents using such attribute processing strategies pose a problem for neoclassical analysis as they cannot be represented by a
conventional utility function (Lancsar and Louviere, 2006). Without continuity,
there is no trade-off between two different attributes (e.g., McIntosh and Ryan,
2002; Rosenberger et al., 2003; Gowdy and Mayumi, 2001). This is a key issue
because without a trade-off, there is no computable marginal rate of substitution
and, crucially for non-market valuation, no computable relative implicit price.
For these reasons, following the sequence of choice tasks, respondents are
often asked to state the attributes they attended to during the experiment thus
enabling the parameters to be conditioned on the basis of attribute neglect or
consideration. The standard practice (e.g., Hensher, 2008; Hensher et al., 2005;
Campbell et al., 2008) is to restrict the parameters to zero for the attributes
respondents have stated they ignored. This approach ensures that unnecessary
weight is not placed on the attributes ignored by respondents. However, if there
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 2
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
are any discrepancies between what respondents say and what they actually do,
restricting the parameters to zero for the attributes they state to ignore is inappropriate. Using a random parameter framework—where we account for heterogeneity in both preferences and attribute processing strategies—this paper analytically investigates the role of attribute processing strategies and explores whether
the respondent’s self-stated responses are consistent with those picked up by the
model. This is a novel approach which tests the aptness of standard practice for
dealing with self-stated attribute processing strategies used in stated choice studies. Results from the analysis provides evidence of significant improvements in
goodness-of-fit and suggests that the standard approach for dealing with attribute
neglect may be inappropriate. The paper uses data from a study that was used
to elicit the economic benefits associated with restoring environmental damage
caused by illegal dumping. Our study focuses on an area close to Belfast, where
illegal dumping activities are prevalent.
The paper is organized as follows. Section 2 defines attribute processing
strategies and discusses possible causes. Section 3 outlines the empirical application, followed by Section 4 in which the random parameters logit models used
in the analysis are presented. Section 5 presents the relevant results. Finally,
Section 6 provides a discussion and offers a number of conclusions.
2
Attribute processing strategies
The estimation of discrete choice experiments assumes that every individual evaluates each and every attribute when choosing their preferred alternative. This implies passive bounded rationality in which individuals are capable of processing
all available information (Puckett and Hensher, 2008). The concept of passive
bounded rationality recognizes that the disparity in attention that individuals allocate to particular attributes is a consequence of the perceived cost and benefits
associated with information evaluation and the opportunity cost of their attention
(DeShazo and Fermo, 2004). Thus a model based on passive bounded rationality
assumes that when individuals are presented with complicated choice sets they
will continue to evaluate all the information provided, however they are more
likely to make mistakes when processing the information (DeShazo and Fermo,
2004).
This typical assumption lends itself to the continuity axiom, which is based
on the notion of unlimited substitutability between attributes. Specifically, individuals are assumed to consider—and make trade-offs—between all attributes
within the choice set. However, recent survey evidence (e.g., Rosenberger et al.,
2003; DeShazo and Fermo, 2002; Sælensminde, 2001; Gelso and Peterson, 2005)
suggests that many respondents exhibit signs of having discontinuous preference
structures (i.e., rationally adaptive behaviour). Attribute processing strategies
imply non-compensatory decision-making behaviour such as lexicographic or-
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 3
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
dering and prevent the marginal rate of substitution between attributes being estimated. In such cases, respondents have a tendency to rank alternatives solely
with reference to a subset of attributes, ignoring all other differences between the
alternatives. Such orderings can be classified according to either ‘strict’ lexicographic procedures—where respondents have an absolute order of preferences
which precludes any degree of substitution between attributes—or ‘modified’
lexicographic preferences—where choice is based on thresholds and minimum
levels of an attribute are necessary (e.g., Lockwood, 1996; Scott, 2002).
Literature suggests a number of reasons as to why individuals might adopt
attribute processing strategies. Attribute processing strategies are likely to be an
indication that there are some attributes within the choice set that are not behaviourally relevant to certain respondents (Sælensminde, 2006). In particular,
these respondents are indifferent with respect to the attributes in the choice set
which they ignore. However, the literature has identified that there is a range of
other factors that may give rise to discontinuous preferences in discrete choice
experiments. The choice tasks respondents are expected to perform require a
significant cognitive effort. Hence, respondents may be unclear how to trade one
attribute against another, and this may well be exacerbated in the case of complex and unfamiliar ecological and environmental goods. Individuals may, therefore, be inclined to impose confines when making trade-offs between attributes
to ease the task. Indeed, a common procedure for some respondents is to consistently discriminate between the attribute(s) they perceive to be more important
and those they perceive to be less important (e.g., Luce et al., 2000; Blamey et al.,
2002; Caussade et al., 2005). Other strategies include establishing thresholds on
attribute levels, conditioning one attribute on the levels of others or choosing
to ignore a subset of the attributes (Hensher et al., 2005). Furthermore, it is
thought that the decision for individuals to employ a coping strategy to deal with
the volume of information they are presented with is intrinsically related to the
perceived complexity of the task. As choice complexity increases—identified in
terms of the number of attributes, the number of choice sets, the number of levels,
the ranges of the attributes and the presentation format—respondents may further
restrict the range of factors that they consider and their precision in evaluation
decreases (e.g., Heiner, 1983; DeShazo and Fermo, 2002; Hensher, 2006; Puckett and Hensher, 2008). As complexity increases, and decision making becomes
more difficult, it is premised that noise is added to the variation in the choices
made by respondents and as a result estimation outputs are derived with less certainty (Puckett and Hensher, 2008). Evidence provided by Arentze et al. (2003)
also confirms that complexity has significant effects on data quality. Therefore, it
is misleading to assume that all individuals have an unlimited capacity to process
information to make a utility-maximizing choice. Subsequently it is important to
recognize how receptive individual respondents are towards complex information
and their inclination to use attribute processing strategies (Hensher et al., 2005).
There is also a range of external factors which may explain the use of heuristics.
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 4
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
These are discussed in Payne et al. (1993) and Rosenberger et al. (2003), and include the cognitive ability of the respondent, the strength of attitudes, beliefs, or
dispositions that the respondent holds, demographic characteristics of the respondent, and the social and economic environment and situation (e.g., distractions
and time pressures during the experiment).
The existence of attribute processing strategies has significant repercussions
for both the design of choice tasks and their estimation. In cases where attribute
processing strategies are not accounted for it is inherently assumed that all the
choice tasks were within the cognitive ability of the respondents, that all attributes were relevant to respondents and the design was of sufficient complexity
so that the choice experiment was meaningful (Hensher et al., 2005). However,
DeShazo and Fermo (2004) brings the validity of this assumption into question.
They argue that if rationally adaptive behaviour is evident it is necessary to account for such to mitigate the significant misspecification bias that would be
present if the model was estimated under the assumption of passive bounded rationality. Thus, it becomes necessary to condition the choice made by individuals
on the information that they claim influenced their choice rather than assuming
attendance to all information.
As a result, analysts are required to identify how respondents evaluate information and the consideration they give to each attribute and acknowledge that
respondents may only consider a subset of the information provided (DeShazo
and Fermo, 2004). Therefore, when designing choice sets it is necessary to deliberate the type and quantity of information which will be considered by respondents to produce an adequately extensive set of choice tasks which will permit
the identification of the attribute processing strategy adopted. From this it can be
established if attribute neglect is employed as a coping strategy or as a process
of evaluating alternatives (Hensher et al., 2005). It is important to note, however,
that if a respondent chooses to ignore a particular attribute this does not necessarily imply that the actual marginal disutility is zero, but instead the cost of fully
considering that attribute may be perceived to outweigh the benefits (Hensher
et al., 2005; Campbell et al., 2008).
By better understanding how respondents attend to information within choice
tasks, there is potential to greatly improve the design of choice models. Furthermore, from an econometric perspective there are obvious benefits in estimating
choice models which condition the choices only on the basis of information that
actually influences respondent’s choices rather than assuming attendance to all
information (DeShazo and Fermo, 2004). Welfare estimates are also likely to
be biased under modelling specifications that neither assume nor allow for violations of the continuity axiom and attribute processing strategies. Therefore,
evidence strongly advocates the use of models which have the capacity to accommodate violations of the continuity axiom and limit potential bias which could
lead to subsequent inaccurate policy implications (Puckett and Hensher, 2008).
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 5
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
For instance, when comparing a baseline model, which assumes passive bounded
rationality with a rationally adaptive model, DeShazo and Fermo (2004) and
Campbell et al. (2008) demonstrate significant differences in willingness to pay
(WTP) estimates derived from the passive bounded rationality assumption. Such
findings indicate important policy implications for models based on the standard
passive bounded rationality assumption as they can be distinctly different from
those derived from a rationally adaptive model which conditions parameter estimates on the attribute processing strategies adopted by respondents (Puckett and
Hensher, 2008). Swait (2001) also argues that accommodating such strategies
can considerably improve the ability of the analyst to predict behavioral changes
associated with proposed policy changes.
3
Empirical application
The introduction of the EU Environmental Liability Directive (2004/35/CE) establishes a common framework for the prevention and restoration of environmental damage. Using discrete choice experiments, this research aimed to examine
public preferences for environmental restoration activities as permitted by the
Directive. The Belfast Hills, where environmental damage arising from illegal
dumping is prevalent, is used as a case study.
The discrete choice experiment exercise reported here involved several rounds
of design and testing. This process began with the gathering of opinions from
stakeholders. Having identified the initial attributes, a series of focus group discussions with members of the public were held. The aims of the focus group discussions were fourfold: to highlight the criteria and issues that the general public
felt were of importance to the countryside surrounding Belfast; to produce and
refine a list of interpretable attributes, and levels thereof, that could later be used
in a discrete choice experiment survey; to shed light on the best way to introduce and explain the choice tasks; and, finally, to provide a platform to test draft
versions of the questionnaire. Following the focus group discussions, the questionnaire was piloted. This pilot testing had the objective of checking whether
the wording and format of the questionnaire was appropriate and if respondents
were able to understand the discrete choice experiment exercises.
In the final version of the questionnaire four attributes were decided upon
to describe the restoration activities. Restorative attributes were categorised as
improvements that could take place at the illegal dump sites or general improvements that could take place elsewhere within the Belfast Hills boundary. This
distinction was made to coincide with the EU Environmental Liability Directive
which stipulates remediation to take place either at the damaged site (i.e., on site)
or at an alternative location geographically linked to the damaged site (i.e., offsite). The discrete choice experiment contained one on-site restoration attribute:
improvement at the Dump Sites, and three complementary, or off-site, restoration
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 6
Attribute processing strategies in stated choice analysis
Danny Campbell and Victoria S. Lorimer
attributes: improvement to Water Quality, Wildlife Habitats and Outdoor Recreation. For each restoration attribute, three possible levels of improvement were
available. To lessen the cognitive burden on the respondent, these levels were
consistent for each attribute. They were described as A Lot of Improvement,
Some Improvement and No Improvement. Each of which was explained in terms
of the level of improvement that would be achieved through their implementation. The Cost attribute was described as a one-off cost (in Pounds Sterling) that
the respondent would personally have to pay to implement the alternative. An
orthogonal design was used to generate a panel of six repeated choice tasks. For
each choice task respondents were asked firstly to indicate their preferred alternative between two experimentally designed alternatives—labelled OPTION A
and OPTION B. Secondly, respondents were asked to choose between their chosen alternative and a DO NOTHING option—which portrayed all the restoration
attributes at the No Improvement level with zero cost to the respondent. An
example of a choice task presented to respondents during the discrete choice experiment is given in Figure 1.
When making their choices, respondents were asked to consider only the attributes presented in the choice task and to treat each choice task independently.
In an attempt to minimize hypothetical bias, respondents were also reminded to
take into account whether they thought restoring the environmental damage was
worth the payment asked of them and were made aware that environmental protection is embedded in an array of substitute and complementary goods. In total,
3234 observables were obtained from a random sample of 556 respondents.
4
Empirical models
In this paper, we use a random parameters logit model specification to account
for unobserved taste heterogeneity. Random parameters logit models provide a
flexible and computationally practical econometric method, which, as described
in McFadden and Train (2000), may in principle be used to approximate any
discrete choice model derived from random utility maximization. Starting with
the conventional specification of utility, we have:
Uni = β0n xni + ni ,
(1)
where Uni is the utility that respondent n obtains from alternative i; βn is a vector
of parameters of variables for respondent n representing the respondent’s tastes;
xni is a vector of observed explanatory variables that relate to alternative i and to
respondent n; and, ni is a Gumbel-distributed and independently and identically
distributed (iid) random term, with constant variance π2 /6, and where we have
assumed a linear in parameters specification of the observed utility function.
A treatment of repeated choices, with preferences varying across respondents,
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 7
Attribute processing strategies in stated choice analysis
Danny Campbell and Victoria S. Lorimer
Figure 1: Example of a choice task presented to respondents
OPTION A
OPTION B
SOME improvement at
dumping sites
A LOT of improvement at
dumping sites
A LOT of improvement in
water quality
NO improvement in
water quality
A LOT of improvement in
wildlife habitats
SOME improvement in
wildlife habitats
SOME improvement in
outdoor recreation
NO improvement in
outdoor recreation
Cost to you
£10
Cost to you
£20
I prefer OPTION A
[ ]
I prefer OPTION B
[ ]
X
X
If instead, there was a ‘DO NOTHING’ option available—which would mean nothing would be done
to deal with the unauthorised dumping and would cost you nothing extra—would the ‘DO NOTHING’
option be your most preferred?
Yes
[ ]
No
[ ]
but not across observations from the same respondent, can also be accommodated. In this case, we work with a sequence of choices for each individual and
treat to be respondent-specific, thus addressing the intrinsic correlation among
observations from the same respondent. Denoting the respondent’s chosen alternative in choice occasion t as ynt and their sequence of choices over the T n choice
occasions as yn = yn1 , yn2 , . . . , ynTn , then, conditional on βn , the probability of
respondent n’s sequence of choices is the product of logit formulas:
Tn
Y
exp (Vnit (βn ))
P (yn |βn , xn ) =
,
J
P
t=1
exp Vn jt (βn )
(2)
j=1
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 8
Attribute processing strategies in stated choice analysis
Danny Campbell and Victoria S. Lorimer
where Vnit = β0n xnit . As βn is not given, the unconditional choice probability
becomes the integral of the logit probability, L (yn |βn , xn ), over all values of βn ,
weighted by the density of βn , f (β):
P (yn |βn , xn ) =
Z
L (yn |βn , xn ) f (βn ) dβn .
(3)
βn
A further advantage of random parameters logit models—which we exploit in
this paper—is that they also accommodate the estimation of respondent-specific
distributions of preferences by deriving the conditional distribution based (within
sample) on their known choices (i.e., prior knowledge) (e.g., Train, 2003; Hensher and Greene, 2003; Sillano and Ortúzar, 2005; Hensher et al., 2006). These
conditional parameter estimates are strictly same-choice-specific parameters, or
the mean of the parameters of the sub-sample of respondents who, when faced
with the same choice tasks, made the choices. Using Bayes’ rule, the conditional
probability is given by Equation 4 (Hensher and Greene, 2003):
h (βn |yn , xn , θ) =
L (yn |βn , xn ) g (βn |θ)
,
P (yn |xn , θ)
(4)
where L (yn |βn , xn ) is now the likelihood of an individual’s sequence of choices if
they had this specific βn ; θ are the parameters of this distribution; g (βn |θ) is the
distribution in the population of βn s; and, Pnit (yn |xn , θ) is the choice probability
defined as:
Z
Pnit (yn |xn , θ) =
L (yn |βn , xn ) g (βn |θ) dβn .
(5)
βn
In this paper such probabilities are approximated in estimation by simulating
the log-likelihood with 500 pseudo-random draws. A key element of the random
parameters logit model is the assumption regarding the distribution of each of
the random parameters. After evaluating the results from various specifications
and distributional assumptions, in estimation we impose each of the K random
parameters to be normally distributed with mean βnk and standard deviation σnk .
Estimation of discrete choice models generally assume respondents are fully
rational, fully informed and behave in utility-maximizing manner. However, as
the previous discussion highlighted, it has become increasingly recognized that
actual respondent behaviour may be somewhat different. Despite this realization,
almost all estimation of discrete choice experiment applications fail to accommodate non-compensatory behaviour. Therefore, as part of the debriefing, we asked
respondents a series of questions that would help identify the attribute processing
strategy they adopted during the discrete choice experiment. This information
can be used within the econometric model to account for the heterogeneity in attribute processing. The standard approach in previous literature is to specify the
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 9
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
attribute parameters as a function of a dummy variable representing whether or
not the attribute was said to be considered (Hensher et al., 2005; Campbell et al.,
2008). In so doing, the choice probabilities are constructed in such a way that
the actual elements of βn that enter the likelihood function are set to zero in cases
where the element is associated with an attribute that was reported to be ignored
by respondent n. However, this approach is quite stringent and does not account
for the fact that there may be some discrepancies between the attribute processing
strategies that respondents said they adopted and those they actually did adopt.
We, therefore, introduce a further model where we do not place any restrictions
on the parameters for attributes reported to be ignored. Essentially this leads to
a model with separate attribute parameters estimated for respondents who said
they considered and ignored the attribute. This provides a convenient approach
for assessing the accuracy of the self-stated attribute processing strategies (i.e.,
if the attribute parameter for respondents who stated they ignored the attribute is
found to be significantly different from zero it implies that the attribute was not
completely ignored by these respondents and, thus, the standard approach may
not best fit the data). We compare this model against the standard approach for
accommodating attribute processing strategies and a model that assumes passive
bounded rationality.
5
5.1
Results
Incidence of discontinuous preferences
Subsequent to the discrete choice experiment, respondents were asked to state
whether they considered or ignored each of the attributes. These self-stated attribute processing strategies are summarized in Table 1. As may be seen, 141
(25%) respondents stated they considered all attributes in the discrete choice experiment. Inspection of Table 1 reveals that 8 (1%) respondents said they ignored
all attributes and a further 13 (2%) said they focused solely on only one attribute,
thus providing no information on their willingness to make trade-offs among the
attributes. When reaching their decisions 74 (13%) respondents indicated that
they took into account two attributes. Three and four attributes were said to be
considered by 155 (28%) and 165 (30%) respondents respectively. With 488
(88%) respondents, Water Quality is the attribute reported to be most considered
by respondents. Overall, 454 (82%), 473 (85%) and 324 (58%) respondents said
they considered the Dump Sites, Wildlife Habitats and Outdoor Recreation attributes respectively. The Cost attribute was said to be considered by 252 (45%)
respondents. Accordingly, the Cost attribute is the attribute that was indicated
to be least taken into account in the discrete choice experiment, which is an
important finding in a study that was primarily concerned with deriving WTP
estimates. This result would suggest that the Cost attribute was the least relevant
factor in influencing the respondent’s choices. Further scrutiny of Table 1 reveals
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 10
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
Table 1: Self-stated attribute processing strategies
Attributes and combinations of attributes considered
Dump Sites, Water Quality, Wildlife Habitats, Outdoor Recreation and Cost
Dump Sites, Water Quality, Wildlife Habitats and Outdoor Recreation
Dump Sites, Water Quality, Wildlife Habitats and Cost
Dump Sites, Water Quality, Outdoor Recreation and Cost
Dump Sites, Wildlife Habitats, Outdoor Recreation and Cost
Water Quality, Wildlife Habitats, Outdoor Recreation and Cost
Dump Sites, Water Quality and Wildlife Habitats
Dump Sites, Water Quality and Outdoor Recreation
Dump Sites, Water Quality and Cost
Dump Sites, Wildlife Habitats and Outdoor Recreation
Dump Sites, Wildlife Habitats and Cost
Dump Sites, Outdoor Recreation and Cost
Water Quality, Wildlife Habitats and Outdoor Recreation
Water Quality, Wildlife Habitats and Cost
Water Quality, Outdoor Recreation and Cost
Wildlife Habitats, Outdoor Recreation and Cost
Dump Sites and Water Quality
Dump Sites and Wildlife Habitats
Dump Sites and Outdoor Recreation
Dump Sites and Cost
Water Quality and Wildlife Habitats
Water Quality and Outdoor Recreation
Water Quality and Cost
Wildlife Habitats and Outdoor Recreation
Wildlife Habitats and Cost
Outdoor Recreation and Cost
Dump Sites
Water Quality
Wildlife Habitats
Outdoor Recreation
Cost
None
Total
Number
Percent
141
113
33
6
3
10
76
8
13
12
4
2
21
16
2
1
18
10
1
9
23
2
4
2
5
0
5
2
3
0
3
8
556
25.36
20.32
5.94
1.08
0.54
1.80
13.67
1.44
2.34
2.16
0.72
0.36
3.78
2.88
0.36
0.18
3.24
1.80
0.18
1.62
4.14
0.36
0.72
0.36
0.90
0.00
0.90
0.36
0.54
0.00
0.54
1.44
100.00
that only 249 (45%) respondents said they made trade-offs between at least one
of the restoration attributes and the Cost attribute.
5.2
Estimation results
Reported in Table 2 are the parameter estimates for three models. Model 1 pertains to the estimation of the data assuming full attribute attention (i.e., passive
bounded rationality). Models 2 and 3 account for the heterogeneity in respondent’s self-stated attribute processing strategies by allowing the parameters to
take different values based on whether or not the attribute was said to be considered. Whilst the attribute parameters for respondents who said they ignored the
attribute are constrained to zero under Model 2 (i.e., the standard approach), they
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 11
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
b
a
2.833
2.277
2.323
0.796
−0.034
critical value equal to 18.31 χ210,0.05 .
critical value equal to 31.41 χ220,0.05 .
L (0)
L β̂
χ2
ρ̄2
AIC
BIC
Considered
Dump Sites
Water Quality
Wildlife Habitats
Outdoor Recreation
Cost
Ignored
Dump Sites
Water Quality
Wildlife Habitats
Outdoor Recreation
Cost
β̂
2.093
1.729
1.970
1.287
0.107
σ̂
−3552.91
−1773.23
3559.36a
0.500
3566.46
3627.27
14.018
13.220
13.071
6.619
−4.417
t-ratio
Model 1
10.367
9.171
9.692
6.576
9.485
t-ratio
Fixed
Fixed
Fixed
Fixed
Fixed
0.000
0.000
0.000
0.000
0.000
1.818
1.506
1.451
1.114
0.102
σ̂
−3552.91
−1750.31
3605.20a
0.507
3520.62
3581.44
16.829
15.414
15.551
8.113
−5.145
t-ratio
2.780
2.169
2.262
1.131
−0.052
β̂
Model 2
Table 2: Parameter estimates
10.245
8.955
8.193
6.309
8.164
t-ratio
1.330
0.945
0.495
0.185
−0.004
3.223
2.514
2.664
1.244
−0.066
β̂
0.690
0.560
0.894
0.979
0.070
2.111
1.796
1.863
1.257
0.131
σ̂
−3552.91
−1678.83
3748.16b
0.526
3397.67
3519.30
6.631
4.137
2.201
1.243
−0.534
14.519
13.330
13.274
7.774
−5.218
t-ratio
Model 3
1.636
1.284
2.408
3.958
6.142
10.100
8.745
8.557
5.310
8.027
t-ratio
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
Page 12
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
are freely estimateable under Model 3. The restoration attributes included in the
models are the two levels (i.e., A Lot of Improvement and Some Improvement)
specified as dummy (1,0) variables, with parameters constrained equal relative to
No Improvement. Although this does not differentiate the effects of the degree
of policy action on the given restoration attribute, it has the advantage of being
parsimonious and sufficient for the purpose at hand, which is assessing whether
or not respondent’s self-stated attribute processing strategies are consistent with
the strategies they actually adopted. To allow heterogeneous preferences among
respondents for all attributes, they are all specified as random with normal distributions.
Under Model 1, the restoration attributes are statistically significant, with positive signs—implying that respondents, all else held constant, prefer environmental damage caused by illegal dumping to be restored (relative to the no improvement condition)—and the Cost parameter is negative, and significant, which is
in line with a priori expectations. The standard deviations are also significant—
indicating that there is heterogeneity in respondent’s tastes. Turning to Model
2, we find that the parameters estimated for the subsets of respondents who said
they considered the attributes are significant and have the expected signs. While
the standard deviations are also significant, we observe that they are markedly
lower (in relative terms) compared to those estimated in Model 1. This finding
complies with the finding in Campbell (2008) and suggests that attribute processing strategies play an important role in the unobserved heterogeneity.
Similarly, under Model 3, the parameters estimated for respondents who said
they considered the attributes are significant and have the expected signs. The
standard deviations are significant, and of the same relative magnitude to those
derived in Model 2. Of greatest interest in Model 3, are the parameters and standard deviations estimated for respondents who said they ignored the attributes.
We remark that these are not zero and are estimated with the expected signs.
Moreover, further inspection reveals that, with the exception of Outdoor Recreation and Cost the estimated parameters are significantly different from zero.
This is an important finding, because for these attributes it suggests that, on the
whole, respondents who stated they ignored these attributes did not completely
ignore them. Thus, fixing the parameters for these respondents to zero is inappropriate. We also note the surprising fact that there is significant heterogeneity
in taste intensities among respondents who said they ignored attributes, for three
of the attributes, namely Wildlife Habitats, Outdoor Recreation and Cost.
Contrasting the parameters estimated for respondents who said they ignored
the attributes against those for respondents who stated they considered the attributes, we find, as expected, that they are lower in the case of the restoration
attributes and higher in the case of the Cost attribute. Further analysis leads to the
rejection of the null that they are equal. A further expected finding, as revealed
by the standard deviations (in relative terms), is that there is substantially less
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 13
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
heterogeneity among respondents who said they ignored the attributes vis-à-vis
those who said they considered the attributes.
While all models are found to be statistically
significant and have acceptable
ρ̄ values, as reflected by the increases in L β̂ and ρ̄2 and reductions in the
AIC and BIC statistics, there is an overall increase in model performance as one
moves from Model 1 to Model 3—indicating that failing to accommodate noncompensatory behavior in discrete choice analysis may lead to biased estimates.
2
To further highlight the features of the model assumptions we report in Figure 2 the box-plots of the conditional distributions derived from Equation 4. The
box-plots show the median, notches to indicate the 95% confidence interval of
the median, ‘hinges’ corresponding with the first and the third quartile (i.e., the
25th and the 75th percentile points in the cumulative distribution), limits of the
plots (using the standard multiple of 1.5 times the interquartile range) and the
resulting outliers. The box-plots show quite clearly that the conditional distributions are different under each of our modelling assumptions. In line with observations from Table 2, the box-plots confirm that, with the sole exception of
the Cost attribute (Figure 2(e)), parameter estimates for respondents who stated
they ignored an attribute are (on average) greater than zero. The box-plots also
illustrate that these distributions also have less spread and variability compared
to those produced by respondents who sated they considered the attribute. Moreover, non-overlapping notches reflect rejection of the null that the median values
for respondents who said they considered an attribute are equal to respondents
who said they ignored an attribute.
An alternative way of teasing out the effect of axiomatic violations of compensatory decision-making, which is likely to be of greater interest to policy makers
is to consider the effects on the WTP estimates. Table 3 reports the marginal
WTP estimates (using the point estimates given in Table 2). Similar to Campbell
et al. (2008), we find stark differences in the WTP estimates derived from the
model that assumes passive bounded rationality (i.e., Model 1) to those derived
when this assumption is relaxed (i.e., Models 2 and 3). Whilst the implied rank
remained constant, we find that the WTP estimates obtained from respondents
who said they considered both the restoration and Cost attributes are much lower
than those where it is assumed that respondents considered all attributes. We
also note that the WTP estimates for these respondents are slightly lower under
Model 3 compared to Model 2. Using Model 3, separate WTP estimates are
derived to identify differences between the four possible self-stated attribute processing strategies: (i) restoration attribute and Cost considered, (ii) restoration
attribute considered but Cost ignored, (iii) restoration attribute ignored but Cost
considered, and (iv) restoration attribute and Cost attribute ignored. Irrespective of whether or not respondents said they considered restoration attributes, we
find that very high WTP estimates—which are a direct consequence of the lowly
estimated denominator (i.e., the Cost parameter) in the WTP calculation—are
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 14
Attribute processing strategies in stated choice analysis
Danny Campbell and Victoria S. Lorimer
Figure 2: Conditional distributions of β̂n
Model 1
Model 2 (considered)
●● ●
Model 3 (considered)
●
Model 3 (ignored)
●
●●
0
2
4
6
(a) Dump Sites
Model 1
●
Model 2 (considered)
●
Model 3 (considered)
●
●
Model 3 (ignored)
●
−2
−1
0
1
2
3
4
5
(b) Water Quality
Model 1
●●
Model 2 (considered)
Model 3 (considered)
●●
●
●
●
●●
Model 3 (ignored)
●
●
●
●
●●
●
●
−1
0
1
2
3
4
5
(c) Wildlife Habitats
Model 1
Model 2 (considered)
●
●●
●
●● ● ●
●●
●●●
Model 3 (considered)
●
Model 3 (ignored)
●
●
●●
●●
●●
●●
● ● ●●
● ● ●●
●● ●
●
●●
●
● ●●
●●●
●●●
●●●●●
●
●
●
●● ●● ●●●
●●●●●
● ●
●●
● ●
−1
●
0
●●
●●
1
2
3
(d) Outdoor Recreation
Model 1
●
●●●●●●●
●
Model 2 (considered)
Model 3 (considered)
●
●
Model 3 (ignored)
● ●●
−0.3
−0.2
● ●● ● ●●●
−0.1
●
0.0
●●
0.1
(e) Cost
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 15
Attribute processing strategies in stated choice analysis
Danny Campbell and Victoria S. Lorimer
Table 3: Willingness to pay estimates (once-off payment in Pounds Sterling)
Model 1
β̂
t-ratio
Model 2
β̂
t-ratio
Considered restoration attribute and considered Cost attribute
Dump Sites
83.41 4.576
53.65 5.237
Water Quality
67.05 4.520
41.86 5.198
Wildlife Habitats
68.40 4.515
43.66 5.164
Outdoor Recreation 23.43 4.010
21.82 4.621
Considered restoration attribute and ignored Cost attribute
Dump Sites
Water Quality
Wildlife Habitats
Outdoor Recreation
Ignored restoration attribute and considered Cost attribute
Dump Sites
Water Quality
Wildlife Habitats
Outdoor Recreation
Ignored restoration attribute and ignored Cost attribute
Dump Sites
Water Quality
Wildlife Habitats
Outdoor Recreation
Weighted averagea
Dump Sites
83.41 4.576
20.00 5.237
Water Quality
67.05 4.520
16.64 5.198
Wildlife Habitats
68.40 4.515
16.43 5.164
Outdoor Recreation 23.43 4.010
6.36 4.621
a
Model 3
Percent of
β̂
t-ratio
respondents
49.19
38.37
40.66
18.99
5.309
5.280
5.300
4.664
37.95
40.47
38.31
29.68
722.60
563.70
597.25
278.94
0.535
0.535
0.534
0.534
43.71
47.30
46.76
28.60
20.35
14.43
7.55
2.82
4.324
3.358
2.065
1.218
7.37
4.86
7.01
15.65
298.91
211.95
110.90
41.42
0.534
0.532
0.520
0.500
10.97
7.37
7.91
26.08
20.17
16.23
16.10
5.63
5.324
5.294
5.305
4.664
100.00
100.00
100.00
100.00
Based on attribute processing strategies adopted by respondents. Insignificant WTP estimates treated as zero.
obtained from respondents who said they ignored the Cost attribute. These estimates exceed what we expect a member of the Belfast public would be willing
to pay to restore environmental damage caused by illegal dumping. Fortunately,
none of these WTP estimates are found to be significant. Interestingly, with
the exception of the Outdoor Recreation attribute, we find that respondents who
stated they ignored the restoration attribute but considered the Cost attribute have
a significant WTP value. We note that the remaining WTP estimates for these
respondents are not as high as those who said they consider both the restoration
and Cost attributes, which is an expected finding.
In Table 3 we also report the sample weighted average WTP estimates—
weighted by the percentage (given in the final column) of respondents who said
they adopted the attribute processing strategy. In the case of Model 1, where it
is assumed that all attributes were attended to by all respondents, no weighting
is necessary. However, for Models 2 and 3, based on the self-stated processing
strategies, less than half of the respondents made a trade-off between the Cost
attribute and any of the restoration attributes, and thus the WTP estimates need
to be weighted accordingly. The weighted WTP estimates derived under Model 3
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 16
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
also account for the proportions of respondents who said they ignored the Dump
Sites, Water Quality and Wildlife Habitats attributes but said they considered
the Cost attribute. Since the remaining self-stated attribute processing strategies
produced WTP estimates that were not significantly different from zero, they
were assigned zero in the weighted average calculation. Whilst we find that the
weighted WTP estimates obtained under Models 2 and 3 are comparable, they
are all approximately 75% lower than those derived under Model 1.
6
Discussion and conclusions
This study was designed to provide straightforward insight into the public’s preferences to restore environmental damage caused by illegal dumping within the
Belfast Hills and, as addressed in this paper, the implications of attribute processing strategies. This was achieved using random parameters logit models,
where we accounted for heterogeneity in both preferences and attribute processing strategies.
Results from this analysis signified that respondents do not attend to all attributes within discrete choice experiments and are likely to adopt an attribute
processing strategy to ease their decision-making. Moreover, from a modelling
point of view, compared to models which accounted for the self-stated attribute
processing strategies, parameters obtained from the base model were found to
be erroneous and biased. As a result, significant improvements in model performance and, thus, more accurate utility expressions were achieved when such
strategies were accounted for in estimation. However, the fact that improvements
in model fit were observed when the parameters for attributes said to be ignored
were freely estimateable, suggests that the standard approach for accommodating attribute processing strategies may be inappropriate. Furthermore, we find
that most of these parameters are significantly different from zero (with significant heterogeneity in tastes), implying that there is some discrepancy between
the self-stated responses and the attribute processing strategies picked up by the
model. This provides further evidence that the standard approach does not adequately deal with the heterogeneity in processing strategies. In relation to the
welfare estimates, in this empirical application, we do not find any substantial
differences between the two approaches for accommodating self-stated attribute
processing strategies. However, we do find that the naı̈ve model, which does not
account for the heterogeneity in attribute processing, results in an over estimation
of the WTP values—in the order of three times the magnitude.
Deciding whether or not to account for attribute processing strategies is a
judgment that should not be based on statistical criteria alone. However, such
strategies do not satisfy the underlying continuity axiom and are a departure from
the use of compensatory decision-making. The fact that a significant proportion
of respondents state they use these simple decision-making heuristics, combined
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 17
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
with the reported effect that their underlying preferences are found to be significantly different and have a major bearing on WTP estimates, suggests some
caution when this issue is neglected in estimating discrete choice models. The evidence presented herein provides compelling evidence for further research in this
area. Future studies should incorporate procedures for identifying and dealing
with attribute processing strategies so that the sensitivity on model performance
and welfare estimates can be further evaluated.
References
Alpı́zar, F., Carlsson, F., Martinsson, P., 2003. Using choice experiments for non-market valuation. Economic Issues 8 (1), 83–109.
Arentze, T., Borgers, A., Timmermans, H., DelMistro, R., 2003. Transport stated choice responses: effects of task complexity, presentation format and literacy. Transportation Research
Part E 39, 229–244.
Bhat, C. R., 2001. Quasi-random maximum simulated likelihood estimation of the mixed multinomial logit model. Transportation Research Part B 35 (7), 677–693.
Bhat, C. R., 2003. Simulation estimation of mixed discrete choice models using randomized and
scrambled Halton sequences. Transportation Research Part B 37 (9), 837–855.
Blamey, R. K., Bennett, J. W., Louviere, J. J., Morrison, M. D., Rolfe, J. C., 2002. Attribute
causality in environmental choice modelling. Environmental and Resource Economics 23,
167–186.
Campbell, D., June 2008. Identification and analysis of discontinuous preferences in discrete
choice experiments. In: European Association of Environmental and Resource Economists
16th Annual Conference. Gothenburg.
Campbell, D., Hutchinson, W. G., Scarpa, R., 2008. Incorporating discontinuous preferences
into the analysis of discrete choice experiments. Environmental and Resource Economics,
401–417.
Caussade, S., Ortúzar, J. D., Rizzi, L. I., Hensher, D. A., 2005. Assessing the influence of design
dimensions on stated choice experiment estimates. Transportation Research Part B 39 (7),
621–640.
DeShazo, J. R., Fermo, G., 2002. Designing choice sets for stated preference methods: the effects
of complexity on choice consistency. Journal of Environmental Economics and Management
2002, 123–143.
DeShazo, J. R., Fermo, G., 2004. Implications of rationally-adaptive pre-choice behaviour for
the design and estimation of choice models. University of California, Los Angeles.
Gelso, B. R., Peterson, J. M., 2005. The influence of ethical attitudes on the demand for environmental recreation: incorporating lexicographic preferences. Ecological Economics 53 (1),
35–45.
Gowdy, J. M., Mayumi, K., 2001. Reformulating the foundations of consumer choice theory and
environmental valuation. Ecological Economics 39, 223–237.
Heiner, R. A., 1983. The origin of predictable behavior. The American Economic Review 73,
560–595.
Hensher, D. A., 2006. Towards a practical method to establish comparable values of travel time
savings from stated choice experiments with differing design dimensions. Transportation Research Part A 40 (10), 829–840.
Hensher, D. A., 2008. Joint estimation of process and outcome in choice experiments and implications for willingness to pay. Journal of Transport Economics and Policy 42 (2), 297–322.
Hensher, D. A., Greene, W. H., 2003. The mixed logit model: the state of practice. Transportation
30, 133–176.
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 18
Danny Campbell and Victoria S. Lorimer
Attribute processing strategies in stated choice analysis
Hensher, D. A., Greene, W. H., Rose, J. M., 2006. Deriving willingness-to-pay estimates of travel
time savings from individual-based parameters. Environment and Planning A 38, 2365–2376.
Hensher, D. A., Rose, J. M., Greene, W. H., 2005. The implications on willingness to pay of
respondents ignoring specific attributes. Transportation 32, 203–222.
Lancaster, K. J., 1966. A new approach to consumer theory. The Journal of Political Economy
74 (2), 132–157.
Lancsar, E., Louviere, J. J., 2006. Deleting ‘irrational’ responses from discrete choice experiments: a case of investigating or imposing preferences? Health Economics 15, 797–811.
Lockwood, M., 1996. Non-compensatory preference structures in non-market valuation of natural area policy. Australian Journal of Agricultural Economics 40, 73–87.
Louviere, J. J., Hensher, D. A., 1982. Design and analysis of simulated choice or allocation
experiments in travel choice modeling. Transportation Research Record 890, 11–17.
Louviere, J. J., Hensher, D. A., Swait, J. D., 2000. Stated choice methods: analysis and application. Press Syndicate of the University of Cambridge, Cambridge.
Louviere, J. J., Woodworth, G., 1983. Design and analysis of simulated consumer choice or
allocation experiments: an approach based on aggregate data. Journal of Marketing Research
20, 350–357.
Luce, M. F., Payne, J. W., Bettman, J. R., 2000. Coping with unfavorable attribute values in
choice. Organizational Behavior and Human Decision Processes 81, 274–299.
McFadden, D. L., Train, K. E., 2000. Mixed MNL models for discrete response. Journal of
Applied Econometrics 15, 447–470.
McIntosh, E., Ryan, M., 2002. Using discrete choice experiments to derive welfare estimates
for the provision of elective surgery: implications of discontinuous preferences. Journal of
Economic Psychology 23, 367–382.
Payne, J. W., Bettman, J. R., Johnson, E. J., 1993. The adaptive decision maker. Cambridge
University Press, Cambridge.
Puckett, S. M., Hensher, D. A., 2008. The role of attribute processing strategies in estimating the
preferences of road freight stakeholders. Transportation Research Part E 44, 379–395.
Rekola, M., 2003. Lexicographic preferences in contingent valuation: a theoretical framework
with illustrations. Land Economics 79, 277–291.
Rosenberger, R. S., Peterson, G. L., Clarke, A., Brown, T. C., 2003. Measuring dispositions for
lexicographic preferences of environmental goods: integrating economics, psychology and
ethics. Ecological Economics 44, 63–76.
Sælensminde, K., 2001. Inconsistent choices in stated choice data: use of the logit scaling approach to handle resulting variance increases. Transportation 28, 269–296.
Sælensminde, K., 2002. The impact of choice inconsistencies in stated choice studies. Environmental and Resource Economics 23, 403–420.
Sælensminde, K., 2006. Causes and consequences of lexicographic choices in stated choice studies. Ecological Economics 59, 331–340.
Scott, A., 2002. Identifying and analysing dominant preferences in discrete choice experiments:
an application in health care. Journal of Economic Psychology 23, 383–398.
Sillano, M., Ortúzar, J. D., 2005. Willingness-to-pay estimation with mixed logit models: some
new evidence. Environment and Planning A 37, 525–550.
Spash, C. L., 2000. Ecosystems, contingent valuation and ethics: the case of wetland recreation.
Ecological Economics 34, 195–215.
Sugden, J., 2005. Anomalies and stated preference techniques: a framework for a discussion of
coping strategies. Environmental and Resource Economics 32, 1–12.
Swait, J. D., 2001. A non-compensatory choice model incorporating attribute cutoffs. Transportation Research Part B 35 (10), 903–928.
Train, K. E., 2003. Discrete choice methods with simulation. Press Syndicate of the University
of Cambridge, Cambridge.
European Association of Environmental and Resource Economists Annual Conference, Amsterdam, June 2009
Page 19