Effects of payment vehicle non-attendance in choice experiments on value estimates and the WTA–WTP disparity

ABSTRACT

In this paper, we argue that value estimates obtained from choice experiments suffer from hypothetical bias, caused by part of the respondents ignoring the payment vehicle in making their choices. We show that this particular form of non-attendance can be substantial, pushes down the estimated payment vehicle parameter, and causes an upward bias in value estimates. Moreover, payment vehicle non-attendance affects willingness to accept (WTA) more than it does willingness to pay (WTP). As a consequence, the WTA–WTP disparity decreases when non-attendance is accounted for, with disparities decreasing by at least 50%. The patterns in findings are by and large robust to exclusion of systematic status quo choices and to alternative model specifications.

KEYWORDS:

• Hypothetical bias
• environmental valuation
• value estimates
• payment vehicle
• attribute non-attendance
• WTA–WTP disparity

JEL:

• C25, Q51

1. Introduction

An important critique of stated preference methods is that they suffer from hypothetical bias, i.e. the notion that individuals behave and choose differently in hypothetical situations than in reality (e.g. Murphy et al. 2005; Hensher 2010). Not only are value estimates from stated preference methods potentially biased and inaccurate, it is also detrimental to the credibility of these methods and as a result to the use of these methods in environmental policy design. Moreover, differences between willingness to accept (WTA) and willingness to pay (WTP) estimates continue to differ widely. In a meta-analysis of WTA–WTP ratios, Horowitz and McConnell (2002) show that large WTA–WTP disparities are widely spread, and moreover that they are prevalent across study subjects and study design, providing evidence that WTA–WTP differences are not just experimental artefacts. An especially worrying result in the context of valuation of public goods, which is often done using hypothetical non-market valuation methods, is that the WTA–WTP ratio is substantially higher for hypothetical experiments than for real experiments (see also List 2003), and for public non-market goods than for private market goods. These observations are especially problematic for environmental valuation, since it often has to rely on stated preference approaches. Moreover, the WTP is theoretically inferior in valuing benefit loss, and, therefore, the WTA is a more logical measure of value in many situations (e.g. Mitchell and Carson 1989; Knetsch 2010).

A (partial) explanation for the observed WTA–WTP disparities is that people value losses more than gains, which has come to be known as loss aversion (Kahneman and Tversky 1979). This hypothesis has been tested in various fields, such as marketing (e.g. Wertenbroch, Soman, and Chattopadhyay 2007), transport economics (e.g. Li and Hensher 2011), health economics (e.g. Chilton et al. 2012; Viscusi and Huber 2012) and environmental economics (Mansfield 1999; Koetse and Brouwer 2016). Another set of explanations are provided by the literature on the effects of bounded rationality, decision heuristics and decision anomalies, on choices made in hypothetical markets (for an overview, see Leong and Hensher 2012). For example, a study by Bateman et al. (2009) shows that using virtual reality instead of the standard representations of choice attributes in choice experiments may reduce the WTA–WTP gap substantially. They show that advanced disclosure of choice attributes and levels reduces respondent judgment error and moderates reliance on loss aversion in uncertain situations. There are several other studies that provide evidence that experience and training mitigate or even fully eradicate the observed WTA–WTP disparity (e.g. List 2003; Plott and Zeiler 2005; Kling, List, and Zhao 2013).

In this paper, we focus on non-attendance in choice experiments as a possible explanation for both hypothetical bias in value estimates and for the substantial disparities between WTA and WTP found in stated preference valuation studies. More specifically, we analyse whether accounting for non-attendance to the payment vehicle in choice experiments (taxes, costs, etc.) reduces hypothetical bias and decreases the disparity between WTA and WTP. The central reasoning and argument in this paper is twofold. First, in choice experiments people may ignore certain attributes, but it is difficult to assess whether this is due to non-attendance or due to non-importance (Hess et al. 2013). Although additional questions can be included in a survey that are aimed at separating non-importance and non-attendance, such questions are likely not easy to understand and difficult to answer for respondents. In the end, non-attendance and non-importance are difficult to separate for most attributes. We argue that the situation is quite different for the payment vehicle. Specifically, it is unlikely that people will ignore changes in prices or taxes in making choices in reality, implying that ignoring such changes in stated preference studies is a form of hypothetical behaviour that may cause bias and that should be corrected for.

Second, in non-market valuation of environmental goods and services, the payment vehicle used is often some form of tax, water or energy bill or donation. Generally, WTP estimates are obtained by presenting people with increases in payments, while WTA estimates are obtained by presenting people with decreases in payments (compensations). The reasons for ignoring the payment vehicle in choice experiments may vary widely (e.g. Saelensminde 2006), but one of the most important arguments is likely consequentiality (e.g. Carson and Groves 2007; Vossler, Doyon, and Rondeau 2012). That is, the reason for ignoring changes in taxes or prices in choice experiments may be strongly related to the notion that the outcomes of the experiment will not lead to such changes in reality. If non-consequentiality is indeed the underlying reason for payment vehicle non-attendance, we may furthermore observe a difference in non-attendance between WTP and WTA experiments. The reasoning is that the occurrence of increases in the taxes or prices may be perceived as credible, but that this is less the case for decreases in taxes or prices, because they (are perceived to) occur less in reality.

When an individual does not attend to the payment vehicle, this basically means a zero tax coefficient for that individual, which artificially decreases the overall payment vehicle parameter estimate and artificially increases welfare estimates (see also Scarpa et al. 2009). Higher non-attendance for tax decreases than for tax increases, therefore, gives rise to and/or increases the WTA–WTP disparity (see next section for more details). Our two main hypotheses are, therefore:

(1)	controlling for payment vehicle non-attendance decreases value estimates obtained from choice experiments and
(2)	controlling for payment vehicle non-attendance decreases the disparity between WTA and WTP.

We test these hypotheses by comparing results of standard choice models with those of attribute non-attendance models, using data from three choice experiments that contain both WTP and WTA choice questions.

The idea that non-attendance to the payment vehicle may lead to inflated value estimates is not new. Scarpa et al. (2009) already show that non-attendance to the cost attribute in their choice experiment leads to increased value estimates. We use a similar but more simplified set-up in which we only address payment vehicle non-attendance, and ignore non-attendance to the non-monetary attributes because of the non-importance problem discussed above. Our contribution to the literature is threefold. First, we argue explicitly that non-attendance to the payment vehicle is a form of hypothetical behaviour, and that controlling for this particular type of non-attendance may reduce hypothetical bias in value estimates. Second, we use data from three separate choice experiments to test this hypothesis, thereby contributing to the stock of empirical evidence on this topic. Third, we explicitly test whether hypothetical bias is larger for WTA than for WTP, and whether it as a result increases the observed WTA–WTP disparity in choice experiments.

The remainder of this paper as organised as follows. In Section 2, we discuss the state-of-the-art in modelling attribute non-attendance. Section 3 discusses the set-up of the choice experiments. In Section 4, we present and discuss estimation results for the three experiments. In Section 5, we discuss the effects of accounting for payment vehicle non-attendance on value estimates and the WTA–WTP disparity. This section is aimed at testing our main hypothesis. Section 6 concludes.

2. Payment vehicle non-attendance in choice experiments

In the rational behavioural model, it is assumed that individuals use all information available to them in making their choices. In the rationally adaptive behavioural model, it is assumed that individuals know that their information processing abilities are limited, and that they process the information such that cognition costs are minimised and benefits of information processing are maximised (DeShazo and Fermo 2004). Based on this adaptive behavioural model, an approach was developed that has become known as attribute non-attendance. Initial applications of this approach are provided in Hensher (2007), Campbell, Hensher and Scarpa (2011) and Hensher, Rose, and Greene (2012). The approach assumes that individuals pay an unequal amount of attention to attributes used in a choice experiment. They pay more attention to attributes that are considered to be more important, and much less attention to or even ignore attributes that are considered to be less important. The reasons for this behaviour may vary from time pressure to cognitive overload to attribute credibility (e.g. Saelensminde 2006; Hensher 2007; Hensher, Rose, and Greene 2012). When not accounted for, the most important problem of attribute non-attendance is that welfare estimates are biased (Campbell, Hutchinson and Scarpa 2008; Puckett and Hensher 2008).

Empirical studies show that accounting for attribute non-attendance in choice models results in a much better fit, but evidence on its effects on welfare estimates is somewhat mixed. While most studies find a substantial decrease in welfare estimates (Campbell, Hutchinson, and Scarpa 2008; Scarpa et al. 2012; Hensher and Greene 2010; Scarpa et al. 2009; Puckett and Hensher 2008), some studies find an increase (e.g. Hensher, Rose, and Bertoia 2007). While these studies look at non-attendance to all attributes, this paper focuses specifically at non-attendance to the payment vehicle. The reason for this is that non-attendance to non-monetary attributes can often be confounded with non-importance (see Hess et al. 2013), while this is arguably not the case for the payment vehicle. The argument is that it is difficult to imagine that individuals would ignore changes in tax, prices or donations in reality, implying that ignoring these changes in a choice experiment is a form of hypothetical behaviour that may cause bias.

To show what happens under payment vehicle non-attendance, note that the value estimate V (WTP or WTA) associated with an attribute or attribute level a is given by(1) where β_a is the parameter estimate on non-monetary attribute a and β_m is the parameter estimate on payment vehicle m. Now, respondents that do not attend to the payment vehicle ignore variation in the payment vehicle, implying that β_m = 0 and that as a consequence V_a goes to infinity for these respondents. Clearly, including these respondents in welfare estimate calculations invariably pushes the value estimate upwards, be it WTA or WTP (see also Scarpa et al. 2009). In conclusion, not accounting for this type of hypothetical behaviour in choice models causes an upward bias in WTP and WTA welfare estimates. Clearly, when non-attendance to decreases in a payment vehicle is higher, for example because decreases are perceived to be less consequential or less credible than increases, upward bias in WTA estimates is higher than for WTP estimates.

3. Models and estimation strategy

For the analyses in this paper, we use an approach in which non-attendance is inferred from the choice data. We use this approach rather than using stated non-attendance because it produces more robust results and because stated non-attendance faces several problems that are related to erroneous reporting by respondents (Carlsson, Kataria, and Lampi 2010; Hess and Hensher 2010; Scarpa et al. 2012; Alemu et al. 2013). Specifically, we estimate an equality-constrained latent class (ECLC) model (see Campbell et al. 2010; Scarpa et al. 2012). The ECLC model uses a two-step procedure. The first step is estimating a latent class model, with the restrictions that ignored attributes have coefficients equal to zero and that coefficients for attended attributes are the same across all classes. These restrictions ensure that the estimated class probabilities actually reflect attribute non-attendance rather than preference heterogeneity between classes. At the second step, either the estimated coefficients are used to calculate welfare values (Hensher, Rose, and Greene 2012), or the estimated class probabilities are used to weight the attribute coefficients in the multinomial logit (MNL) model so as to account for individual degrees of attribute non-attendance (Campbell et al. 2010; Scarpa et al. 2012).

For all experiments, we estimate and compare the outcomes of two models. First, we estimate a classic MNL model, in which the probability of choosing alternative i among j = 1, …, J alternatives is given by(2) where x is a vector with non-monetary choice attributes, m represents the payment vehicle, and β and δ represent parameters to be estimated by the model.

The second model is an ECLC model that incorporates and corrects for non-attendance to the payment vehicle, while ignoring non-attendance to all other attributes. In this model, the latent classes are not used in the classic sense, i.e. to represent heterogeneity in people's preferences, but are used to model heterogeneity in attribute processing strategies, or in our case non-attendance to the payment vehicle. The reason for choosing this particular ECLC specification, and its strength, is that it quite rigorously controls for what I argue to be hypothetical bias due to payment vehicle non-attendance. Although more complex specifications are possible, for example specifications in which more heterogeneity is allowed for the payment vehicle parameter, these models provide similar patterns but for value estimates that are in-between those from CL and from our ECLC specification (see Section 7.2 for details). The value estimates obtained from our specification should, therefore, be interpreted as lower bound estimates of WTP.

In the ECLC model, we model payment vehicle non-attendance only, implying we include two classes. In the first class, the coefficient on the payment vehicle is estimated by the model. In the second class, this coefficient is restricted to zero. Coefficients for all other attributes are estimated by the model, but are restricted to be equal in both classes, as is required in the ECLC model (Scarpa et al. 2009). Specifically, using the notation in Glenk et al. (2015), the probability of respondent n choosing alternative i among j = 1,…, J, given that there are c = 1 and 2 classes, is given by(3) where θ is a set of two-class-specific constants that are identified by restricting their sum to zero. Note that the model is different from the CL model, but only in that the payment vehicle parameters δ are class specific. Crucial is furthermore that δ₁ is estimated by the model and that δ₂ is restricted to 0 in order to test for payment vehicle non-attendance. An important outcome of the ECLC model is the set of class probabilities, in this case for the 2 classes. These class probabilities can be used to assess the probability that the payment vehicle has been ignored in making a choice between the choice alternatives. The larger is non-attendance to the payment vehicle, the larger is the proportion of respondents in the second class. Conversely, without non-attendance the proportion of respondents belonging to the second class should be zero.

With respect to model estimation, three issues need to be addressed. First, we estimate MNL and ECLC models separately for WTP and WTA data in order to allow for differences in scales between WTP and WTA experiments. See the next section for details on the WTP–WTA data-sets used in this study. Second, our analytical focus is solely on payment vehicle non-attendance. One reason is that although non-attendance to other attributes is potentially relevant, it is likely equally relevant for both WTA and WTP questions and the effects of non-attendance to other attributes cancel out when looking at the WTA–WTP ratio. Moreover, a central argument of this paper is that non-attendance to the payment vehicle is a good measure of hypothetical bias, i.e. it measures non-attendance rather than non-importance because changes in costs, prices and tax are arguably not ignored in reality. For other attributes, this is much more difficult to argue, and non-attendance is confounded with non-importance. Third, in our models, we do not control for preference heterogeneity. Not because we think this is unimportant in general, but because the samples drawn for the WTP and WTA choice experiments are either identical (experiment 1), or were drawn in an identical fashion, implying that differences in WTP and WTA sample composition are small and cannot account for our results.

4. Choice experiment data

The choice experiments used for this study were not designed specifically to test the central hypotheses in this paper. Instead we aimed at collecting data from a diverse set of choice experiments from the literature. The reasons for this were to test our hypotheses in various contexts, and to test them for various payment vehicles. Because we analyse payment vehicle non-attendance and the differential effects on WTP and WTA value estimates, the choice experiments we were looking for needed to include both increases and decreases in the payment vehicle. Well-known choice experiments by Bateman et al. (2009) and De Borger and Fosgerau (2008) contain both WTP and WTA value estimates, but measure these through increases and decreases in non-monetary attributes, and not through increases and decreases in the payment vehicle. These were, therefore, excluded. We ultimately obtained data from three choice experiments. The first two experiments use in- and decreases in local taxes, while the third experiment uses in- and decreases in the water bill, so to some extent we can test our hypotheses for different payment vehicles.¹ Below we present and discuss in detail the three choice experiments used for this study.

4.1. Experiment 1

Data for experiment 1 were obtained from a choice experiment aimed at assessing consumer preferences for different types of natural areas and their characteristics. For this, respondents were presented with choice alternatives that are of a generic nature instead of referring to a specific site or area. From the literature, it is evident that consumer preferences for natural areas may be affected by many characteristics. In this experiment, the attributes are: the type of natural area, its size, the distance from the residence to the area, accessibility of the area, degree of fragmentation and changes in annual municipal tax, which is an annual tax in the Netherlands that is levied separately from national income taxes. In order to be able to estimate both WTP and WTA estimates, we include both tax increases and tax decreases in the experiment, but in separate choice tasks. A summary of attributes and attribute levels is provided in Table 1. For further details on this experiment, we refer to Koetse, Verhoef, and Brander (2016).

Table 1. Attributes and attribute levels in experiment 1.

Attributes	Levels
Type of natural area	▪ Small-scale grassland
	▪ Natural area with water
	▪ Forest
Size of natural area	▪ 2 km^2
	▪ 6 km^2
	▪ 16 km^2
Distance to natural area	▪ 1 km
	▪ 5 km
	▪ 15 km
Fragmentation of natural area by urban sprawl	▪ No fragmentation (one patch)
	▪ Medium fragmentation (two patches)
	▪ High fragmentation (four patches)
Fragmentation of natural area by transport infrastructure	▪ No fragmentation
	▪ Medium fragmentation (one road)
	▪ High fragmentation (two roads)
Accessibility for recreation	▪ Accessible
	▪ Not accessible
Annual municipal tax	▪ Increase/decrease of €320
	▪ Increase/decrease of €140
	▪ Increase/decrease of €60
	▪ Increase/decrease of €20
	▪ No change

Source: Adapted from Koetse, Verhoef, and Brander (2016).

Choice options were presented using both text and figures. These figures represent all attributes except for the payment vehicle, which was included in the choice card below the graph. An opt out or status quo alternative was excluded because the aim was to identify generic relative preferences for natural areas and their characteristics in the Netherlands, rather than to identify whether consumers prefer change to no change in a specific situation. This may affect the absolute value of welfare estimates to some extent, but it should not affect our analysis of payment vehicle non-attendance and its relative consequences for welfare estimates.

A fractional factorial statistical design was generated, containing 100 survey versions of 12 choice tasks each. For data collection, a Dutch internet panel managed by TNS-NIPO was used, containing over 200,000 households. The panel is established through random sampling, meaning that each member of society has an equal chance to be added to the panel as long as he or she has conveyed the willingness to cooperate. Throughout the entire data collection process, respondents were sampled using representative sampling (for the entire Dutch population) on age, gender, education, household size and size of municipality. A total of 2100 questionnaires were sent out, and ultimately 1360 complete responses were obtained, implying a response rate of nearly 65%. For the choice experiment we have a total of 16,102 observations.

4.2. Experiment 2

Data for experiment 2 are obtained from a study carried out around the IJsselmeer, one of the largest freshwater buffers in Europe (for details, we refer to Koetse and Brouwer 2016). This buffer is used during the summer as one of the main sources of water supply for agriculture and residential household water demand. In the winter the lake functions as a buffer for excess storm and river water. In order to anticipate future climate change and corresponding droughts and floods, the Dutch government is considering a future increase in the IJsselmeer water level. In this study different samples were used for WTP and WTA, in contract to experiment 1. The attributes and attribute levels for the WTP and WTA experiments are summarised in Table 2.

Table 2. Attributes and attribute levels experiment 2.

Attributes	Attribute levels
Raising the dikes	▪ No (reference scenario in WTP version)
	▪ Yes (reference scenario in WTA version)
Flood probability	▪ Once every 5000 years (reference scenario WTA)
	▪ Once every 1000 years
	▪ Once every 500 years (reference scenario WTP)
Change in bird population	▪ Zero per cent (reference scenario WTA)
	▪ −10%
	▪ −30% (reference scenario WTP)
Type of shore	▪ No shores (reference scenario WTP)
	▪ Shores off the coast
	▪ Shores next to the dike (reference scenario WTA)
Change in annual local tax (increases for WTP and decreases for WTA)	▪ €0 (reference scenario)
	▪ €60
	▪ €100
	▪ €180

Source: Koetse and Brouwer (2016).

A fractional factorial main effects design was generated, containing 15 versions of 10 choice tasks each. Each respondent was randomly assigned to one of these 15 versions. Although the statistical design for WTP and WTA was identical, the reference situations were different. The reference situation in the WTP experiment is a situation that is most unfavourable with respect to flood probability, presence of shores and bird populations, but in which no tax increases take place. The two alternative policy scenarios in the WTP version are more favourable with respect to at least one of these three non-monetary attributes, accompanied by an increase in annual local taxes. The reference situation in the WTA version is a situation that is most favourable with respect to flood probability, shores and bird populations, but in which there are no tax reductions. The two alternative policy scenarios in the WTA version are less favourable with respect to at least one of the three non-monetary attributes, and are accompanied by compensation in the form of a decrease in annual local tax.

Identical to experiment 1, the sample was drawn from the TNS-NIPO internet panel, but where experiment 1 was implemented across the entire Netherlands, experiment 2 assesses preferences of people who live close to the IJsselmeer. In the TNS-NIPO panel, around 6800 respondents live in the postal code areas around the IJsselmeer. From this set, two independent samples were drawn. For each sample, we employed representative sampling based on age, gender, household size, social class (education and profession) and residential location. For each of the two choice experiments, 375 households were invited to complete the survey. The total number of respondents for the WTP experiment is 298, while for WTA it is 314, yielding an average overall response rate of around 80 per cent. The socio-demographic respondent characteristics of the two samples reveal that the samples are very similar in composition. We exclude respondents that systematically chose the status quo option, just to avoid that such differences between WTP and WTA experiments affect our results. This leaves us with 251 respondents for WTP and 201 respondents for WTA.

4.3. Experiment 3

Data for experiment 3 were obtained from a study on consumer preferences for urban water service standards (MacDonald, Morrison, and Barnes 2010). More specifically, consumers were asked for their preferences regarding water supply interruptions. In the choice experiment the duration of the interruption, the number of interruptions, the way of communicating the interruption, methods of alternative water supply and changes in costs of the water bill were included as attributes. This study is of particular interest since it includes a different payment vehicle (costs of water bill) than the other two studies (local tax), and may, therefore, shed light on whether results are affected by the type of payment vehicle.

Attributes and attribute levels for this choice experiment are summarised in Table 3. As the table shows, attributes and attribute levels are not identical in the WTA and WTP experiments, implying that comparing WTA and WTP estimates for this choice experiment will give unreliable estimates of the WTA–WTP disparity. For example, changes in the costs of the water bill are not identical for WTA and WTP: (1(1) ) WTA questions present a one-time rebate while WTP questions present a change in annual costs and (2(2) ) the amounts of monetary changes are different. This leads to differences in WTA and WTP that are not purely related to loss aversion, but also to non-linear preferences for costs and the difference between annual increases (WTP) and one-time rebate (WTA). Still, we should be able to filter out the effects of payment vehicle non-attendance on WTA and WTP estimates, even though the obtained differences between WTA and WTP are not accurate reflections of loss aversion. Because of the asymmetry of the design, we will only analyse the effects for those attribute levels that are identical for WTA and WTP.

Table 3. Attributes and attribute levels in experiment 3.

Attributes	Levels WTA (first levels represent current practice)	Levels WTP (first levels represent current practice)
Duration of interruption	▪ 3 hours	▪ 6 hours
	▪ 6 hours	▪ 4 hours
	▪ 9 hours	▪ 2 hours
Expected number of interruptions in twelve months	▪ No more times	▪ Two more times
	▪ One more time	▪ One more time
	▪ Two more times	▪ No more times
Communication	▪ Card in your letter box	▪ Card in your letter box
	▪ Radio station announcement	▪ Phone call
	▪ Information only if you phoned in	▪ Knock on your door
Alternative water supply offered	▪ 2 litre bottle of water	▪ Nothing unless you requested it
	▪ Water tanker in a central location	▪ Water tanker in a central location
	▪ Nothing unless you requested it	▪ 2 litre bottle of water
Change in costs of water bill	▪ $0	▪ $0
	▪ $25 off next water bill (one time)	▪ Increase by $40 (annual)
	▪ $50 off next water bill (one time)	▪ Increase by $80 (annual)

Source: Adapted from MacDonald, Morrison, and Barnes (2010).

After excluding all ‘no choice’ answers, there are 1321 observations for WTA and 1111 observations for WTP. The statistical design contained 192 choice cards, equally divided among 32 survey versions, i.e. six choice cards per version. The statistical design was generated such that attribute levels were different for the two non-status-quo options. Although the design was not strictly orthogonal, the correlation between attributes was small and averaged around 0.06.

Similar to the second experiment, different samples of identically sampled individuals were used for the WTP and WTA choice experiments. The study was performed in Adelaide, Australia, and samples for the experiments were drawn from streets that had experienced a water supply interruption in the past. A drop-off, pick-up format was used to give people time to think about their choices. The response rate was around 80%, with 230 respondents for the WTA and 197 respondents for the WTP experiment. Also for this experiment, we exclude respondents that systematically chose the status quo option, to avoid that such differences between WTP and WTA experiments affect our results. This leaves us with 155 respondents for WTP and 135 respondents for WTA.

5. Estimation results

5.1. Estimation results experiment 1

As is the standard in discrete choice models, the choice variable in the three experiments is modelled as a binary dependent variable. For experiment 1 we include size of the area, distance to the area and tax as continuous explanatory variables. Attribute levels for type, fragmentation and accessibility of the natural area are included as dummy coded explanatory variables, with ‘grassland’, ‘no fragmentation’ and ‘accessible areas’ as the reference categories, respectively. MNL and ECLC estimation results for experiment 1 are presented in Table 4.

Table 4. MNL and ECLC estimation results for experiment 1 (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	WTP		WTA
Attribute variables	MNL	ECLC	MNL	ECLC
Water (d)	0.181** (0.040)	0.224** (0.045)	0.274** (0.037)	0.282** (0.039)
Forest (d)	0.614** (0.042)	0.769** (0.047)	0.692** (0.038)	0.767** (0.041)
Size (in km^2)	0.019** (0.003)	0.022** (0.003)	0.021** (0.003)	0.023** (0.003)
Distance (in km)	–0.026** (0.003)	–0.029** (0.003)	–0.034** (0.003)	–0.038** (0.003)
Fragmentation: medium (d)	–0.065 (0.040)	–0.078 (0.045)	–0.050 (0.037)	–0.061 (0.039)
Fragmentation: strong (d)	–0.385** (0.042)	–0.378** (0.047)	–0.416** (0.038)	–0.457** (0.040)
Area not accessible (d)	–0.648** (0.038)	–0.761** (0.045)	–0.745** (0.036)	–0.780** (0.038)
Tax increase/decrease (in 100 Euro)	–0.906** (0.024)	–2.128** (0.117)	0.153** (0.015)	4.629** (0.813)
Prob. class 1 (No ANA)	–	0.691** (0.021)	–	0.117** (0.013)
Prob. class 2 (ANA to tax)	–	0.309** (0.021)	–	0.883** (0.013)
NOBS	8116	8116	7986	7986
Log–L	–4192	–4031	–4944	–4852
Pseudo R2 (adjusted)	0.254	0.283	0.106	0.123

** and * = statistically significant at 1% and 5%, respectively.

Estimation results for the MNL and ECLC model are quite similar. In both models water and forest are preferred to grasslands, size of an area increases and distance to an area decreases its value, medium fragmentation has a small negative effect while strong fragmentation has a substantial negative effect on the value of an area, and also when an area is not accessible for recreation its value decreases substantially (for details see Koetse, Verhoef, and Brander (2016)). The first difference between the two models is that the ECLC model has a higher explanatory power than the MNL model; around 3 percentage points for WTP and around 2 percentage points for WTA. The second difference is that the tax coefficients are substantially higher relative to other attribute coefficients in the ECLC than in the MNL model. Another interesting finding is with respect to the estimated latent class probabilities. The share of respondents that does not attend to tax increases is around 31%, while this share is around 88% for WTA, i.e. around 3 times larger.

5.2. Estimation results experiment 2

For experiment 2 we include flood probability, bird population and tax as continuous explanatory variables. Attribute levels type of shore are included as dummy coded explanatory variables, with ‘no shore’ as the reference category. MNL and ECLC estimation results for experiment 2 are presented in Table 5.

Table 5. MNL and ECLC estimation results for experiment 2 (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	WTP		WTA
Attributes	MNL	ECLC	MNL	ECLC
No increase in dike height (status quo constant)	–1.318** (0.086)	–2.453** (0.129)	–0.290* (0.119)	–0.458** (0.111)
Years in which one flood occurs (in 100 years)	0.009** (0.001)	0.010** (0.002)	0.021** (0.002)	0.025** (0.002)
Shores off the coast (d)	0.279** (0.069)	0.252** (0.073)	0.315** (0.084)	0.391** (0.093)
Shores next to dike (d)	0.402** (0.072)	0.361** (0.077)	0.460** (0.088)	0.606** (0.103)
Bird population (in %)	–0.024** (0.002)	–0.023** (0.003)	–0.029** (0.003)	–0.037** (0.004)
Tax increase/decrease (in Euro)	–0.014** (0.001)	–0.035** (0.001)	0.003** (0.001)	0.022** (0.001)
Prob. class 1 (No ANA)	–	0.569** (0.039)	–	0.234** (0.050)
Prob. class 2 (ANA to tax)	–	0.431** (0.039)	–	0.766** (0.050)
NOBS	2510	2510	2010	2010
Log–L	–2383	–2053	–2002	–1795
Pseudo R2 (adjusted)	0.103	0.254	0.062	0.186

** and * = statistically significant at 1% and 5%, respectively.

All coefficients for both models have the expected signs and are statistically significant at least at 5%. Increasing flood probabilities decrease utility, crating shores along the coast have a positive effect on preferences, and decreases in bird population are valued negatively. Very similar to experiment 1, the ECLC model has a higher explanatory power than the MNL model; around 15 percentage points for WTP and around 13 percentage points for WTA. Also the estimated tax coefficients are substantially higher compared to other attribute coefficients in the ECLC than in the MNL model, and more so for tax decreases than for tax increases. With respect to class probabilities, the share of respondents that does not attend to tax increases is around 45%, while the share for tax decreases is around 70%, i.e. almost 1.5 times larger.

5.3. Estimation results experiment 3

For experiment 3, we include duration of interruption, number of interruptions and water bill costs as continuous explanatory variables. Attribute levels for communication and alternative water supply are included as dummy coded explanatory variables, with ‘card in your letter box’ and ‘2 litre bottle of water’ as the reference categories, respectively. MNL and ECLC estimation results for experiment 3 are presented in Table 6.

Table 6. MNL and ECLC estimation results for experiment 3 (ECLC cost coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	WTP		WTA
Attribute variables	MNL	ECLC	MNL	ECLC
Status quo constant	–0.383 (0.277)	–0.276 (0.328)	0.250 (0.214)	0.136 (0.239)
Duration of interruption (in hours)	0.063 (0.042)	0.109* (0.048)	–0.002 (0.023)	0.008 (0.027)
Number of interruptions	0.417** (0.093)	0.432** (0.111)	–0.234** (0.071)	–0.337** (0.086)
Communication level 1 (d)	0.044 (0.163)	0.023 (0.189)	–0.565** (0.129)	–0.590** (0.154)
Communication level 2 (d)	0.020 (0.187)	0.090 (0.217)	–0.571** (0.146)	–0.840** (0.182)
Supply: 2 litre bottle of water (d)	–0.019 (0.194)	0.036 (0.225)	0.002 (0.153)	–0.006 (0.182)
Supply: water tanker (d)	0.123 (0.191)	0.116 (0.239)	0.336** (0.130)	0.400* (0.156)
Cost increase/decrease (in Australian dollars)	–0.058** (0.003)	–0.078** (0.005)	0.018** (0.003)	0.087** (0.014)
Prob. class 1 (no ANA)	–	0.922** (0.032)	–	0.268** (0.053)
Prob. class 2 (ANA to cost)	–	0.078* (0.032)	–	0.732** (0.053)
NOBS	881	881	782	782
Log–L	–582	–449	–811	–655
Pseudo R2 (adjusted)	0.380	0.534	0.044	0.233

** and * = statistically significant at 1% and 5%, respectively.

Although the estimation results are plausible, a problem is that there are quite a few insignificant coefficients. Next to the asymmetric design, this hinders comparison of value estimates between WTP and WTA and between the MNL and ECLC models. The only attribute with a robust coefficient in terms of magnitude and statistical significance, and one that is comparable between experiments because its levels are identical for WTP and WTA, is the number of interruptions.

Other than that, the patterns that emerge are very similar to those for the previous experiments. The explanatory power of ECLC models is larger than power of MNL models; around 16 percentage points for WTP and around 19 percentage points for WTA. Also the estimated tax coefficients are higher compared to other coefficients in the ECLC than in the MNL model, and more so for tax decreases than for tax increases. With respect to class probabilities, the share of respondents that does not attend to water bill costs is around 8%, while the share for tax decreases is around 73%, i.e. around nine times larger.

6. Value estimates and the WTA/WTP disparity

6.1. Effects of non-attendance on value estimates

In the previous section, we showed that non-attendance to the payment vehicle can be substantial, and appears to be systematically higher for decreases than for increases, i.e. WTA is more affected than WTP. In this section we explore the consequences for value estimates and the WTA–WTP disparity. For each experiment, the WTP and WTA value estimates for the MNL and ECLC models are derived using Equation (1).² Because the WTP and WTA value estimates are not normally distributed, it is difficult to parametrically test the statistical difference between estimates from different models (e.g. using the Wald test). We, therefore, statistically test whether value estimates from the MNL model differ from value estimates from the ECLC model using the random sampling procedure by Poe, Giraud and Loomis (2005, 359). This test basically compares value estimates from the two experiments by (1(1) ) making 1000 draws from the distributions of the parameters, determined by the estimated coefficients and their standard errors, (2(2) ) deriving WTP (WTA) values for the parameters for each of the 1000 draws, (3(3) ) calculating the proportion of WTP (WTA) estimates from one model that is larger than WTP (WTA) estimates from a second model, (4) reshuffling the WTP (WTA) combinations from the two models a hundred times, calculating the proportion described in step 3 in each case, and (5) calculating the average proportion. The WTP value estimates for the three experiments are derived from the model estimates presented and discussed in the previous section. The WTP estimates and the Poe test results are presented in Table 7.³

Table 7. WTP estimates from MNL and ECLC models for the three experiments.

Attributes	WTPMNL	ECLC	Difference	Poe test p-value on difference MNL-ECLC
Experiment 1
Water (d)	€20.0	€10.5	−47%	0.028
Forest (d)	€67.8	€36.2	−47%	0.000
Size (in km^2)	€2.13	€1.03	−52%	0.001
Distance (in km)	−€2.88	−€1.38	−52%	0.000
Fragmentation medium (d)	−€7.15	−€3.64	−49%	0.224
Fragmentation strong (d)	−€42.5	−€17.7	−58%	0.000
Inaccessible (d)	−€71.5	−€35.7	−50%	0.000
Experiment 2
Years in which one flood occurs (in 100 years)	€0.67	€0.30	−56%	0.000
Shores off the coast (d)	€19.8	€7.17	−64%	0.009
Shores next to dike (d)	€28.5	€10.3	−64%	0.000
Bird population (in %)	−€1.72	−€0.66	−62%	0.000
Experiment 3
Number of interruptions	€7.12	€5.54	−22%	0.230

Note: The Poe test p-value represents the p-value of the hypothesis test that MNL value estimate ≠ ECLC value estimate.

The table clearly reveals that WTP estimates decrease substantially when accounting for payment vehicle non-attendance. Specifically, WTP estimates decrease by around 47%–58%, 56%–64% and 22% for experiments 1, 2 and 3, respectively. The Poe, Giraud and Loomis (2005) test results show that for most attribute levels the difference between MNL and ECLC value estimates is statistically significant at 95% or more. There are two exceptions. WTP estimates for ‘fragmentation medium’ in experiment 1 are larger for MNL than for ECLC, but only at 77.6%, while WTP for ‘Number of interruptions’ in experiment 3 is larger for MNL than for ECLC, but only at 77%.

The WTA estimates and Poe test results are presented in Table 8. The table shows that WTA estimates decrease even more than WTP estimates when accounting for payment vehicle non-attendance; estimates decrease by around 96%–97%, 81%–83% and 70% for experiments 1, 2 and 3, respectively. In conclusion, not accounting for non-attendance to the tax attribute in choice experiments may lead to substantial overestimation of value estimates, and more so for WTA than for WTP. Poe, Giraud and Loomis’s (2005) test results show that the differences between MNL and ECLC value estimates are statistically significant at 95% or more, with no exceptions, indicating that WTA estimates drop substantially and significantly when accounting for payment vehicle non-attendance.

Table 8. WTA estimates from MNL and ECLC models for the three experiments.

Attributes	WTAMNL	ECLC	Difference	Poe test p-value on difference MNL-ECLC
Experiment 1
Water (d)	€180	€6.09	−97%	0.000
Forest (d)	€454	€16.6	−96%	0.000
Size (in km^2)	€13.6	€0.50	−96%	0.000
Distance (in km)	−€22.2	−€0.81	−96%	0.000
Fragmentation medium (d)	−€32.5	−€1.31	−96%	0.019
Fragmentation strong (d)	−€273	−€9.86	−96%	0.000
Inaccessible (d)	−€488	−€16.8	−97%	0.000
Experiment 2
Years in which one flood occurs (in 100 years)	€6.67	€1.12	−83%	0.000
Shores off the coast (d)	€99.7	€17.4	−83%	0.002
Shores next to dike (d)	€145.5	€27.0	−81%	0.000
Bird population (in %)	−€9.17	−€1.66	−82%	0.000
Experiment 3
Number of interruptions	€13.13	€3.88	−70%	0.016

Note: The Poe test p-value represents the p-value of the hypothesis test that MNL value estimate ≠ ECLC value estimate.

6.2. Effects of non-attendance on the WTA/WTP disparity

The fact that WTA decreases more than does WTP implies that the WTA/WTP ratio decreases as well when taking payment vehicle non-attendance into account. WTA/WTP ratios for the various attribute levels in the three experiments, and Poe test results on the difference between WTA and WTP from MNL and ECLC models, are presented in Table 9. The ratios obtained from the ECLC model are substantially lower than ratios obtained from the MNL model, even to the extent that WTA is smaller than WTP for experiments 1 and 3. For experiment 1, the Poe test confirms that WTA is larger than WTP in the MNL case, but that WTP is larger than WTA for the ECLC model. For experiment 2, the WTA/WTP ratio is substantially reduced, but Poe test results show that WTA is still larger than WTP when accounting for payment vehicle non-attendance. For experiment 3, results are less conclusive, but because the WTA/WTP ratio for this study is not solely a reflection of loss aversion (see Section 3), it is difficult to assign a weight to this result.

Table 9. WTA and WTP estimates from MNL and ECLC models for the three experiments.

	WTA/WTP ratio		Poe test p-value on difference WTA–WTP
Attributes	MNL	ECLC	MNL	ECLC
Experiment 1
Water (d)	9.0	0.6	0.000	0.055
Forest (d)	6.7	0.5	0.000	0.003
Size (in km^2)	6.4	0.5	0.000	0.011
Distance (in km)	7.7	0.6	0.000	0.024
Fragmentation medium (d)	4.6	0.4	0.103	0.158
Fragmentation strong (d)	6.4	0.6	0.000	0.017
Inaccessible (d)	6.8	0.5	0.000	0.003
Experiment 2
Years in which one flood occurs (in 100 years)	10.0	3.8	0.000	0.000
Shores off the coast (d)	5.0	2.4	0.002	0.013
Shores next to dike (d)	5.1	2.6	0.000	0.000
Bird population (in %)	5.3	2.5	0.000	0.000
Experiment 3
Number of interruptions	1.8	0.7	0.089	0.204

Note: The Poe test p-value represents the p-value of the hypothesis test that WTA ≠ WTP.

The changes in value estimates and WTA/WTP ratios are summarised in Figures 1 and 2. These two figures reveal the two most important conclusions in this study. First, both WTP and WTA value estimates decrease substantially when controlling for payment vehicle non-attendance, and not controlling for payment vehicle non-attendance may produce severe overestimations of value estimates. Second, the effect of payment vehicle non-attendance is larger for WTA than for WTP, and as a result the WTA/WTP ratio decreases substantially after non-attendance is controlled for.

Figure 1. Average change in WTP and WTA value estimates for the three experiments due to controlling for payment vehicle non-attendance.

Figure 2. Average WTA–WTP ratios for the MNL and the ECLC models for the three experiments.

7. Robustness of findings

7.1. Inclusion of systematic status quo choices

The results discussed above were obtained for models from which respondents that systematically chose the status quo option were excluded. In our opinion, this is the correct approach for obtaining insight into the effects of payment vehicle non-attendance, since it ensures that a trade-off between attributes takes place. Still, systematic status choices may reveal actual preferences and should in that case be included in a welfare analysis. We, therefore, also estimate models including respondents with systematic status quo choices. For experiment 1, this does not affect results, since in this experiment no status quo option was provided in the choice cards. For experiments 2 and 3, the share of respondents with systematic quo options is substantially higher for the WTA experiment (36% and 41%, respectively) than for the WTP experiment (16% and 21%, respectively). This in itself may already be an indication that decreases in the payment vehicle are considered to be less consequential and/or credible than increases. The estimation results are provided in Appendix 1, and are very similar to those reported in the previous sections, on most accounts. Payment vehicle non-attendance is substantially higher for WTA than for WTP, WTA and WTP value estimates are substantially lower for the ECLC model than for the MNL model, and the WTA/WTP disparity obtained from the ECLC model decreases for experiment 3.

The only and notable difference is that the WTA/WTP disparity for experiment 2 does not decrease when controlling for payment vehicle non-attendance. For this experiment, the WTA and WTP value estimates obtained from the ELC model decrease by similar amounts, and the disparities for MNL and ECLC are, therefore, almost identical. This is despite the fact that non-attendance to tax decreases remains substantially higher than non-attendance to tax increases. On closer inspection, it appears that the drop in WTP value estimates is much more substantial for the sample with than for the sample without systematic status quo choices, while the drop in WTA value estimates is very similar in both samples. Clearly, our findings are by and large robust to the decision on inclusion or exclusion of systematic status quo choices.

7.2. Alternative ANA-model specifications

As discussed earlier in Section 3, there are various other ECLC model specifications that could have been used. We will not analyse results of model with non-attendance to other attributes than the payment vehicle, since for this type of non-attendance it remains difficult to distinguish between non-attendance and non-importance. In order to assess robustness of our findings to other specifications we estimate ECLC models with more than two classes, allowing for more heterogeneity in the payment vehicle parameter while still maintaining one class with a payment vehicle parameter restricted to zero. We estimate WTP and WTA models for 3 and 4 classes in total, containing, respectively, 2 and 3 classes with varying payment vehicle parameters. Parameters on other attributes are again restricted to be equal across classes. Full estimation results for both 3-class and 4-class ECLC models are presented in Appendix 2. In Table 10, we present class probabilities for the class with the payment vehicle coefficient restricted to zero for the 2-, 3- and 4-class ECLC specifications. In the table, we compare results of the new models (3- and 4-class models) with the original ECLC model (2-class model). The table shows that the probability of belonging to the class with a zero payment vehicle coefficient often becomes smaller for models with more classes. This is intuitive since people with a non-zero but small payment vehicle coefficient are now in a separate class. An exception is the WTA model for experiment 1, for which the class probability does not change.

Table 10. Class probabilities for the class with the payment vehicle coefficient restricted to zero for various ECLC specifications.

	WTP			WTA
	2-class ECLC	3-class ECLC	4-class ECLC	2-class ECLC	3-class ECLC	4-class ECLC
Experiment 1	0.309**	0.143**	0.113**	0.833**	0.883**	n.a.^a
Experiment 2	0.431**	0.313**	0.300**	0.766**	0.687**	0.467**^,^b
Experiment 3	0.078*	0.041	0.015	0.732**	0.476*	0.476*

** and * = statistically significant at 1% and 5%, respectively.

^aThe four class WTA model for experiment 1 did not converge.

^bSmall and statistically insignificant tax coefficients of the wrong sign for classes 2 and 3.

Also quite intuitive is that the WTP and WTA value estimates for the 3- and 4-class models are in-between value estimates from the MNL model and value estimates from the original ECLC model. In most cases, however, the drop in value estimates compared to the MNL model is still substantial (see also notes under Table 10). In conclusion, although qualitatively our findings are robust to other ECLC model specifications, WTP and WTA estimates obtained from our original ECLC model should be interpreted as lower-bound value estimates. The WTA–WTP disparities for the 3- and 4-class ECLC models are similar to those for our original ECLC model; for the 3-class ECLC model, the disparities obtained for experiments 1–3 are 0.28, 2.99 and 1.48, respectively. In conclusion, the WTA–WTP disparities obtained from the various ECLC models are all smaller than those from the MNL model, and the magnitude of change varies somewhat across ECLC specifications.

8. Conclusions and discussion

Using three choice experiments that deal with environmental valuation and that contain both WTP and WTA questions, we show that non-attendance to the payment vehicle in choice experiments is substantial, and pushes down the payment vehicle parameter. Since it is difficult to imagine that people would ignore the payment vehicle in reality, we argue that this is a form of hypothetical bias that should be accounted for, and that environmental value estimates obtained from choice experiments will generally overstate true values when this issue is ignored, confirming our first hypothesis. We also show that non-attendance to accepting a payment (WTA) is larger than to making a payment (WTP), leading to larger decreases in WTA estimates than in WTP estimates. As a result the WTA–WTP disparities found for the three experiments go down substantially, confirming our second hypothesis. The disparities found after correction are far more in line with disparities found for revealed preference experiments (see Horowitz and McConnell 2002), and even suggest that the disparity may be non-existent in some cases. Although our main results suggest that loss aversion disappears for two out of three experiments, results from robustness analyses suggest that loss aversion may persist after controlling for payment vehicle non-attendance, depending mainly on the ECLC specification. This is in line studies like the one by Bateman et al. (2009) in which loss aversion is reduced by improving the research design (virtual reality in this case) but does not dissipate entirely. In this light, a potentially interesting area for further research is aimed at uncovering which factors affect the existence and extent of loss aversion.

Our results appear to hold for different payment vehicles, although the evidence on this particular topic is rather thin and further research is required. Also most of these results are robust to the inclusion of respondents who systematically chose the status quo and to other model specifications, although it is clear that the value estimates obtained from our specific non-attendance model should be interpreted as lower bound estimates. The findings suggest that both increases and decreases in the payment vehicle are considered to be inconsequential by parts of the respondent samples (e.g. Carson and Groves 2007; Vossler, Doyon, and Rondeau 2012), and decreases far more so than increases. The non-attendance model and approach used in this study are a simplified version of the one used in most non-attendance studies (e.g. Scarpa et al. 2009), and provides a relatively straightforward way to control for an important source of hypothetical bias, producing lower bound value estimates from choice experiments.

The findings in this study have several implications for using choice experiments for the valuation of ecosystem services in particular, and public goods in general. First, both WTP and WTA decrease substantially when controlling for non-attendance to the payment vehicle. Non-attendance models could, therefore, be used alongside other, more traditional models in order to obtain more conservative, and in our opinion more reliable value estimates. Second, non-attendance to compensations appears to be much larger than non-attendance to payments, leading to a larger upward bias in WTA than in WTP and increasing the WTA–WTP disparity. Controlling for non-attendance decreases this ratio to values that are much more in line with disparities found in revealed preference studies. Our findings are, therefore, both worrisome and hopeful. They are worrisome because a large part of respondents appears to ignore the monetary attribute, implying that larger samples are needed in order to obtain representative and reliable parameter and value estimates. They are hopeful because the recent development of non-attendance models makes it possible to filter out these ‘anomalies’ in stated choice behaviour. Moreover, since WTA is often a more logical value measure than WTP in valuing benefit loss (e.g. Mitchell and Carson 1989; Knetsch 2010), especially for environmental and public goods and services, our specific non-attendance model may be used to obtain more reliable WTA estimates that can be used in welfare calculations and environmental policy decisions.

Further research may be directed at several issues. First, future studies may be used to assess whether other case studies produce similar results, and confirm our conclusions with respect to the effects of payment vehicle non-attendance on value estimates. This may serve as robustness check, but more importantly it may serve to check whether the results diverge systematically for different payment vehicles. For example, non-attendance to tax changes may be substantially different than non-attendance to changes in price, and as a result changes in value estimates after controlling for non-attendance may be smaller. A conclusion could be that tax as a payment vehicle in environmental valuation is often suitable because the study is concerned with a public good, but that other payment vehicles may be more suitable in terms of improved consequentiality, especially when using the WTA as the measure of value. Although this study provides some evidence on this issue, it is far from conclusive and rather thin in scale in scope.

Second, because large non-attendance rates imply that large samples are needed for obtaining representative and reliable value estimates, it interesting to develop and test ways of reducing payment vehicle non-attendance. Cheap talk scripts may be used for this, although empirical evidence on their effectiveness is mixed (e.g. Aadland and Caplan 2006). ‘Real’ choice experiments, in the sense that a choice made in one of the choice cards is actually carried out, are an interesting option because this should eliminate hypothetical behaviour (e.g. Alfnes et al. 2006), at least with respect to the payment vehicle. However, evidence on the effectiveness of this approach is mixed as well (e.g. Carlsson and Martinsson 2001). Moreover, this option is not always available, especially in environmental and public good valuation, because changes in the payment may be too large or simply impossible to enforce in reality. Developing new ways of enticing people to include the payment vehicle in making their choices may, therefore, be necessary.

Finally, conducting choice experiments that are specifically designed to analyse the impact of payment vehicle non-attendance are required to provide stronger evidence on the issues addressed in this paper. Specifically, non-attendance to a variety of payment vehicles in otherwise identical experiments, and the use of consequential and inconsequential designs, are issues that can be addressed better this way.

Acknowledgements

I thank Darla Hatton MacDonald, Mark Morrison and Mary Barnes for sharing their choice experiment data, Danny Campbell for useful comments and suggestions on an earlier draft of the paper, and an anonymous reviewer for highly constructive comments and suggestions.

Disclosure statement

No potential conflict of interest was reported by the author.

Additional information

Funding

The author acknowledges funding from the European Commission seventh Framework Programme [grant number FP7-ENV-2012-308393-2] (OPERAs, http://www.operas-project.eu/).

Notes

1. Another interesting choice experiment is the one by Hu, Adamowicz, and Veeman (2006), but data used for this study were no longer available due to privacy restrictions.

2. Note that the asymmetry of attribute levels and especially the payment vehicle in experiment 3 implies that we cannot interpret the resulting WTA/WTP disparity as a measure of loss aversion. However, effects of controlling for payment vehicle non-attendance can still be used to test our hypotheses.

3. For experiment 2, we exclude the results on the alternative specific constant (ASC), both for WTP and WTA. Although the pattern for the ASC parameters is the same as for the other attributes, the ASC coefficient itself is very sensitive to the type of model and model specification, and, therefore, the changes in value estimates are quite different than those for the other attributes.

Appendix 1. Estimation results including systematic status quo choices

Table A1.1. MNL and ECLC estimation results for experiment 2, including those respondents that systematically chose the status quo option (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	WTP		WTA
Attributes	MNL	ECLC	MNL	ECLC
No increase in dike height (status quo constant)	−1.234** (0.089)	−2.186** (0.122)	1.947** (0.112)	2.482** (0.072)
Years in which one flood occurs (in 100 years)	0.009** (0.001)	0.007** (0.001)	0.019** (0.002)	0.024** (0.001)
Shores off the coast (d)	0.237** (0.065)	0.139* (0.064)	0.297** (0.080)	0.459** (0.075)
Shores next to dike (d)	0.357** (0.068)	0.157* (0.065)	0.429** (0.083)	0.646** (0.081)
Bird population (in %)	−0.021** (0.002)	−0.014** (0.002)	−0.027** (0.003)	−0.039** (0.003)
Tax increase/decrease (in Euro)	−0.013** (0.001)	−0.046** (0.002)	0.003** (0.001)	0.022** (0.001)
Prob. class 1 (no ANA)	–	0.434** (0.031)	–	0.250** (0.033)
Prob. class 2 (ANA to tax)	–	0.566** (0.031)	–	0.750** (0.033)
NOBS	2980	2980	3140	3140
Log-L	−3011	−2395	−2645	−2237
Pseudo R2 (adjusted)	0.077	0.268	0.044	0.351

** and * = statistically significant at 1% and 5%, respectively.

Table A1.2. MNL and ECLC estimation results for experiment 3, including those respondents that systematically chose the status quo option (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	WTP		WTA
Attribute variables	MNL	ECLC	MNL	ECLC
Status quo constant	0.138 (0.251)	0.275 (0.294)	1.196** (0.192)	1.382** (0.241)
Duration of interruption (in hours)	0.058 (0.038)	0.072 (0.043)	−0.002 (0.021)	0.001 (0.026)
Number of interruptions	0.325** (0.082)	0.346** (0.098)	−0.219** (0.064)	−0.249** (0.080)
Communication level 1 (d)	0.096 (0.145)	0.165 (0.166)	−0.473** (0.118)	−0.557** (0.147)
Communication level 2 (d)	0.067 (0.164)	0.164 (0.184)	−0.426** (0.133)	−0.554** (0.166)
Supply: 2 litre bottle of water (d)	−0.089 (0.176)	−0.075 (0.201)	−0.053 (0.141)	−0.064 (0.195)
Supply: water tanker (d)	−0.070 (0.172)	−0.101 (0.210)	0.132 (0.118)	0.242 (0.151)
Cost increase/decrease (in Australian dollars)	−0.053** (0.003)	−0.069** (0.005)	0.015** (0.003)	0.062** (0.005)
Prob. class 1 (no ANA)	–	0.921** (0.031)	–	0.295** (0.079)
Prob. class 2 (ANA to cost)	–	0.079* (0.031)	–	0.705** (0.079)
NOBS	1111	1111	1321	1321
Log-L	−758	−623	−1,167	−911
Pseudo R2 (adjusted)	0.309	0.487	0.024	0.370

** and * = statistically significant at 1% and 5%, respectively.

Appendix 2. Estimation results 3-class and 4-class ECLC models

Table A2.1. 3-class and 4-class ECLC estimation results for experiment 1 (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	3-class model		4-class model
Attribute variables	WTP	WTA	WTP	WTA^a
Water (d)	0.202** (0.046)	0.282** (0.034)	0.208** (0.047)	–
Forest (d)	0.779** (0.049)	0.767** (0.036)	0.796** (0.050)	–
Size (in km^2)	0.023** (0.003)	0.023** (0.003)	0.023** (0.003)	–
Distance (in km)	−0.029** (0.003)	−0.038** (0.003)	−0.030** (0.004)	–
Fragmentation: medium (d)	−0.106* (0.047)	−0.061 (0.041)	−0.099* (0.047)	–
Fragmentation: strong (d)	−0.455** (0.049)	−0.457** (0.040)	−0.451** (0.050)	–
Area not accessible (d)	−0.795** (0.048)	−0.780** (0.035)	−0.805** (0.049)	–
Tax increase/decrease (in 100 Euro): class 1	−7.262** (1.261)	4.629* (1.897)	−26.58 (47.90)	–
Tax increase/decrease (in 100 Euro): class 2	−0.890** (0094)	1.897 (6.7^17)	−3.022** (0.731)	–
Tax increase/decrease (in 100 Euro): class 3	–	–	−0.696** (0.108)	–
Prob. class 1 (no ANA)	0.298** (0.032)	0.117 (0.102)	0.149** (0.039)	–
Prob. class 2 (no ANA)	0.559** (0.032)	0.000 (0.094)	0.261** (0.048)	–
Prob. class 3 (no ANA)	–	–	0.477** (0.048)	–
Prob. class 4 (ANA to tax)	0.143** (0.028)	0.883** (0.019)	0.113** (0.033)	–
NOBS	8116	7986	8116	–
Log-L	−3940	−4852	−3932	–
Pseudo R2 (adjusted)	0.299	0.122	0.300	–

** and * = statistically significant at 1% and 5%, respectively.

^aModel did not converge.

Table A2.2. 3-class and 4-class ECLC estimation results for experiment 2 (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	3-class model		4-class model
Attributes	WTP	WTA	WTP	WTA
No increase in dike height (status quo constant)	−3.960** (0.158)	1.564** (0.080)	−4.162** (0.172)	0.370** (0.090)
Years in which one flood occurs (in 100 years)	0.013** (0.002)	0.027** (0.001)	0.013** (0.002)	0.027** (0.002)
Shores off the coast (d)	0.323** (0.079)	0.461** (0.077)	0.336** (0.079)	0.315** (0.074)
Shores next to dike (d)	0.490** (0.081)	0.676** (0.084)	0.499** (0.082)	0.511** (0.082)
Bird population (in %)	−0.029** (0.003)	−0.041** (0.003)	−0.030** (0.003)	−0.035** (0.003)
Tax increase/decrease (in Euro): class 1	−0.063** (0.002)	0.087** (0.010)	−0.080** (0.004)	0.081** (0.009)
Tax increase/decrease (in Euro): class 2	−0.026** (0.002)	0.015** (0.001)	−0.052** (0.003)	−0.022 (0.012)
Tax increase/decrease (in Euro): class 3	–	–	−0.026** (0.002)	−0.013 (0.009)
Prob. class 1 (no ANA)	0.264** (0.033)	0.083** (0.020)	0.131** (0.031)	0.085** (0.020)
Prob. class 2 (no ANA)	0.423** (0.038)	0.230** (0.048)	0.193** (0.034)	0.191 (0.420)
Prob. class 3 (no ANA)	–	–	0.376** (0.041)	0.258 (0.375)
Prob. class 4 (ANA to tax)	0.313** (0.035)	0.687** (0.051)	0.300** (0.034)	0.467** (0.084)
NOBS	2510	2010	2510	2010
Log-L	−1887	−1745	−1867	−1733
Pseudo R2 (adjusted)	0.315	0.208	0.321	0.213

** and * = statistically significant at 1% and 5%, respectively.

Table A2.3. 3-class and 4-class ECLC estimation results for experiment 3 (ECLC tax coefficients that are constrained to zero are not shown, standard errors are in parentheses).

	3-class model		4-class model
Attribute variables	WTP	WTA	WTP	WTA
Status quo constant	−0.265 (0.358)	0.030 (0.266)	−0.270 (0.363)	0.300 (0.266)
Duration of interruption (in hours)	0.131* (0.055)	0.005 (0.028)	0.132* (0.051)	0.005 (0.028)
Number of interruptions	0.468** (0.121)	−0.315** (0.087)	0.468** (0.121)	−0.315** (0.087)
Communication level 1 (d)	−0.015 (0.195)	−0.596** (0.159)	−0.019 (0.201)	−0.596** (0.159)
Communication level 2 (d)	0.111 (0.229)	−0.842** (0.186)	0.110 (0.257)	−0.842** (0.186)
Supply: 2 litre bottle of water (d)	0.009 (0.235)	−0.053 (203)	0.009 (0.240)	−0.053 (0.203)
Supply: water tanker (d)	0.090 (0.284)	0.389* (0.164)	0.088 (0.244)	0.389* (0.164)
Cost increase/decrease (in Australian dollars): class 1	−0.129** (0.046)	0.110** (0.020)	−0.144 (0.117)	0.110** (0.020)
Cost increase/decrease (in Australian Dollars): class 2	−0.045** (0.012)	0.025* (0.010)	−0.051 (0.029)	0.025 (0.018)
Cost increase/decrease (in Australian Dollars): class 3	–	–	−0.015 (0.060)	0.025 (0.018)
Prob. class 1 (no ANA)	0.638** (0.188)	0.205** (0.056)	0.584* (0.276)	0.205** (0.056)
Prob. class 2 (no ANA)	0.321 (0.167)	0.319 (0.183)	0.343 (0.183)	0.157 (1381)
Prob. class 3 (no ANA)	–	–	0.057 (0.097)	0.162 (1381)
Prob. class 4 (ANA to cost)	0.041 (0.034)	0.476* (0.192)	0.015 (0.093)	0.476* (0.192)
NOBS	881	782	881	782
Log-L	−439	−653	−439	−653
Pseudo R2 (adjusted)	0.543	0.235	0.543	0.234

** and * = statistically significant at 1% and 5%, respectively.