This study presents the results of a valuation exercise which uses CE to estimate the benefits of the environmental attributes for a potential coastal program. The respondents are presented with several sets of choices, which are made up of various alternatives. Each alternative is comprised of a combination of several environmental attributes which reflect the improved outcomes. The composition of the attributes and their levels are varied systematically in each individual alternative, according to experimental design theory (Adamowicz et al., 1998).

As mentioned in the Study Area section, according to the main characteristics of the four coastal parks, the experimental design includes five attributes: beach type, garbage, crowding, water quality and cost. Detailed descriptions of the attribute levels and variables are summarized in Table 1. Specifically, the beach types include Cobblestone, Gravel, and Sand based on the real situations in each of the coastal parks. For the garbage and crowding levels, we combined the real and virtual situations. In terms of water quality, this study focuses on the recreational value from the tourists’ perspective, so the water quality is reflected through the visibility depth of the water on the seaside.

All possible choice sets can be generated with a full factorial design. The total number of combinations of the four environment attributes and cost at different levels is 810 (3×3×5×3×6). In a practical sense, such large choice sets cannot be answered by each respondent (i.e., choosing one out of 180 choices). So, the orthogonal design method was adopted to decrease the number of choice sets (Louviere et al., 2000), and it reduced the number of profiles to 25 alternatives. However, the number of alternatives was still considered too large to allow clear and easy consideration by the respondents.

Ultimately, the number of choice sets was determined by following the procedures proposed in Street et al. (2005). Then the choice sets were randomly blocked into 12 versions. Each choice set included two assigned alternatives and a constant status quo alternative, and all levels of the attributes referred to the present situation. The respondents were asked to provide answers regarding their preferences for the different levels of attribute combinations across the alternatives in the choice sets. This meant that in one choice set each respondent had to choose either an improved environmental situation (Alternatives 2 or 3) or the current situation (Status quo-Alternative 1). The current environmental conditions of a study area constituted the unique status quo choice (Alternative 1).

Most tourists have no concept of WTP, and they are also not used to participating in this type of survey. Therefore, to facilitate the respondents’ understanding of the questionnaire, the survey was administrated in a field interview setting, with the assistance of photographs, to describe the hypothetical environmental quality changes. An illustration of one choice set is presented in Fig. 2.

The survey was comprised of three sections. The first section introduced the survey and contained the questions related to visitors’ attitudes towards environmental quality and their preferences for outdoor recreation activities. Other questions included the importance of natural settings and environmental factors, time spent on each activity, and respondents’ attitudes toward the environmental quality perceptions of the beach. The second part asked visitors about the motivations for their park visits. This information included the places they visit and their primary recreational activities. The final section included questions concerning respondents’ socio-economic status or demographics (i.e., age, gender, income, education, and residence).

The survey was designed in cooperation with Dalian tourism management personnel. A pilot test was conducted on small groups of visitors at Fujiazhuang Park and Xinghai Beach Park during the week of May 24-31, 2019. The final questionnaires were fine-tuned and re-tested during the focus visitor groups. The field survey took place at the end of July 2019, and lasted for ten days. Four sites were selected for the data collection. These sites cover the most important coastal recreational areas in Dalian: Golden Pebble Beach, Fujiazhuang Park, Xinghai Beach, and Xinghai Bay Artificial Beach (Fig. 1). These four sites accommodate the vast majority of the tourists who come to Dalian every year, accounting for approximately 90% of the annual visitation trips. This popularity is why these four sites are under extremely high environmental stress, especially during the peak of the tourism seasons.

The field survey was carried out by the investigators at each beach site. Questionnaires were distributed to the visitors who expressed a willingness to participate in the survey when intercepted by the investigators. Each survey form was filled out with the interviewer’s assistance, which ensured better response rates and questionnaire completeness. The final dataset was comprised of 630 completed questionnaires, collected from 733 interviews. The demographic results are presented in Appendix 1.

The analysis of the marginal utility of the environmental attribute changes encountered some problems since only the Crowding and Cost variables were set up as continuous variables in the survey questions. All the remaining attribute variables were set up as dummy variables to reflect the situation of the attribute as being either present or not present (Table 1).

CE have been widely applied in the fields of environmental management and evaluation (Nie et al., 2019), and reveal consumer preferences using the random utility approach (McFadden et al., 2000). Also, CE provide respondents with hypothetical change conditions while considering a decision process under multiple environmental attributes, as opposed to a single change through considering a consumer’s utility as a composite of the utilities originating from the multiple underlying characteristics of the goods consumed. Given that environmental services contain a number of underlying characteristics, CE provides a useful instrument for valuing environmental goods with multiple attributes and characteristics (Hanley et al., 2001).

According to the random utility theory, the representative individual is assumed to have a utility function that takes the following form:

This function can be decomposed into two parts: a deterministic element and a stochastic element. The deterministic element (V) is derived from any option which is assumed to depend on the attributes (X) (e.g., water quality, crowding of the surrounding, level of vegetation coverage) (Mendoza-González et al., 2018). At the same time, different socioeconomic characteristics (S) could also influence the level of utility (U). The socioeconomic characteristics are typically specified as a linear index of the attributes of the many j alternatives in a choice set. The stochastic element (e) considers the researcher’s unobservable effects on individual choices. Hence, Equation (1) can be re-written as:

The probability that any particular respondent prefers option k in the choice set A over any alternative option i, can be expressed as the probability (Prob) that the utility associated with option k exceeds the utility associated with the other options. Thus:

Data obtained from the questionnaires were analyzed using the conditional logit (CL), heteroscedastic logit (HL), and random parameter logit (RPL) models. The CL is a basic model in CE studies. CL has a typical assumption of independent and identical distribution, with the Type 1 extreme-value distribution. Hence, it offers a good starting point and benchmark for an analysis. However, the CL also assumes that the error variance is constant, usually being set to 1 across individuals. This assumption has been called into question, especially when comparing utility parameter vectors across two datasets. This is problematic because this assumption might mislead the hypothesis testing and cannot be separately identified from the vector of the utility parameter β (Louviere et al., 2002).

Therefore, in this study we also tested HL as an alternative model, as suggested by DeShazo et al. (2002). Since CL is limited by the Independence of Irrelevant Alternatives (IIA), it fails to account for the unobserved heterogeneity and potential correlations between the available choices. Given this limitation, more flexible approaches are desirable.

In this study, we also used the RPL model as another alternative for analyzing the preferences associated with the establishment of coastal tourism environmental programs. The RPL generalizes the CL by allowing the coefficients to vary randomly across individuals (Train, 1998). Therefore, RPL mitigates IIA and can represent any substitution pattern. Furthermore, the RPL explicitly accounts for unobserved heterogeneity (Carlsson et al., 2003).

Table 2 shows that the estimates of the utility parameters vary among the three models. According to the AIC criteria, the HCL and RPL models are both better than the CL model; and this is also evident by the results of the likelihood ratio (LR) test. The LR test shows a model’s fitness with the data. The CL is nested within the RPL, and the RPL is statistically preferred over the CL at the 0.01 significance level (2ΔLL=32.36>χ²(7) =18.48). Furthermore, the RPL results reveal that there is a significant preference heterogeneity, as the estimated standard deviations of both levels of the crowding (Crowding) and high-water quality (WaterQ_high) attributes are statistically significant (P<0.05).

Beach type (BT_Sand) shows statistical significance (P< 0.1). This result could be partly attributed to the well-known problem of heteroscedasticity, which causes the coefficient estimates in discrete choice models to be inconsistent (Yatchew and Griliches, 1985). Furthermore, it also reveals a behavioral interest in factors which influence the variance of the latent variables in the model (Louviere, 2001).

It is possible to compare the CL and HCL models, since the CL is nested within the HCL. Interestingly, the HCL is statistically preferred over the CL at the 0.01 level of significance (2ΔLL=30.12> χ²(5) =15.09). This manifests the presence of a discrete preference heterogeneity. The individual income variable has a significant positive effect on the visitor’s decision-making process. In contrast to the CL-HCL relationship, the HCL is not nested within the RPL. Therefore, it is not necessary to make a comparison test for these two models. However, the HCL agrees with the RPL, suggesting that a significant preference heterogeneity exists.

Furthermore, the RPL model can offer more information regarding the visitors’ preferences for the level of crowding and the water quality improvement. However, those preferences are heterogeneous between the groups. Therefore, according to the comparative analysis, we decided to use the RPL model to conduct a more thorough analysis of the tourism environment preferences.

The results of the RPL indicate that most of the estimated parameters are statistically significant (P<0.1), except for the variables of BT_Sand and WaterQ_med. As expected, the estimated cost parameter is negative, which is consistent with economic theory, and the so-called law of demand.

Regardless of whether the heterogeneity is considered, Crowding has a negative effect. Hence, visitors prefer to encounter fewer people on the site. It is interesting to note that the respondents display a lower sensitivity to the level of crowding, even at a higher level of crowding, because it has a lower estimated coefficient and odds radio in the statistical tests^①( ^① The coefficient of the non-linear model cannot be directly interpreted as a marginal impact on the level of individual utility tourist’s decisions. Each odds ratio value is the multiplicative effect of a unit change of a given independent variable and the odds ratio=exp(b_i), where b_i is the coefficient of the models (Train, 2009).). This insensitivity may be due to the fact that the Chinese tourists are more accustomed to congestion in parks than visitors from other countries (since China has the highest population density in the world).

The tourists did not exhibit any differences in their preferences towards the features of sand and gravel characteristics between the beach sites. The remaining attributes are all positive and statistically significant by at least (P<0.1) in the RPL model. Garbage_med (4-8 pieces), Garbage_low (< 2 pieces) and WaterQ_high (Visibility depth 3-5 feet) all have higher estimated coefficient values and odds radio test results.

In summary, the visitors have stronger preferences for a reduction in the amount of garbage and an increase in water quality at the beaches, suggesting that the current levels of water quality and trash management are not satisfactory to the visitors.

None of the socio-economic variables were significant, except for the income variable in the HCL model. This implies that as the consumer’s income rises, they will more likely choose a higher willingness to pay value, while all other factors are held constant. Based on this result, the level of personal income is probably the primary factor that leads to the heterogeneous preferences among the respondents, although other potential factors could also cause the heterogeneous preferences. Therefore, it will be informative to conduct a welfare analysis using the RPL model, by subdividing the sample data into different groups based on the socio-economic criteria.

Although the preceding coefficient estimates could be used to assess the marginal effect-based odds ratios, unfortunately, they are not easily interpreted from a visitor’s point of view. To facilitate their interpretability from the welfare perspective, generic choice experiments often report the marginal WTP (MWTP). The MWTP can be interpreted as the average marginal WTP for a one-level change of each attribute (Train, 1998). Using a simple linear utility function, the marginal rate of substitution between an attribute and money is simply the ratio between the coefficient of that attribute and the marginal utility of money (Parsons and Kealy, 1992):

where β_j is the coefficient of an analyzed attribute and λ is the cost coefficient. This ratio is often known as the implicit price. MWTP shows how much money an individual is willing to sacrifice for a decrease of one unit in the attribute level. In this study, all environmental attributes are designed as non-continuous variables in the models. Thus, strictly speaking, the ratios of the attribute coefficients and the marginal utility of money are not MWTP, since we cannot express a marginal change for a discrete variable. In such a case, the WTP measure is interpreted as the amount of money a respondent is willing to pay for a change in the attribute from, say, not available to available (Holmes and Boyle, 2005).

As for the compensating-variation welfare measurements, the MWTP from the RPL model states that any one of the chosen attributes with a relatively higher-level of improvement is able to generate a positive amount of welfare gain, as shown by the positive utility parameter estimates (Table 3). For instance, the average welfare gain from the factor of the trash amount (Garbage_low, less than two pieces) is larger than that of the other factors, according to the RPL model. This is followed by the garbage amount (Garbage_med, 4-8 pieces) and water quality (WaterQ_high, visibility depth 3-5 feet) variables.

Crowding has a smaller positive welfare gain resulting from the rate of the crowding level decrease. For example, 12 yuan will be gained as the congestion level is reduced to 100 people. This result is consistent with the previous analysis. However, the beach type (BT_Gravel) shows a unique significant negative welfare result (-17.3 yuan), which is worse than the base alternative of beach type (BT_Stone) of 27.2 yuan^②(^② Since we are using effects coded data, the WTP values for the attribute levels of the base alternative can be derived from the two estimated WTP values of the attributes in question as a negative sum of these two values (i.e. WTP-BT_Stone = WTP-BT_Gravel ×-1+WTP-BT_Sand×-1) (Juutinen et al., 2011).). Overall, the welfare analysis results are consistent with those of the former section of the visitor’s preference analysis.

Furthermore, the coefficients for garbage amount (Garbage_med), water quality (WaterQ_high) and the intensity of crowding (Crowding) are all statistically significant at P<0.05, according to the results of RPL model (Table 2). This may suggest that the level of utility is heterogeneous among the various groups of respondents, which is indicated by the group-specific parameters developed by this RPL model. Using the estimated individual parameter values for each random variable, we were able to generate the average marginal WTP for each subgroup of respondents using the information from the respondents’ characteristics (e.g., traveler generating region, types of activities, and travel purposes).

Table 4 reveals some significant differences, in terms of economic values, between the visitor groups. The attributes examined (e.g., traveler generating region, gender, level of education) show no significant effects on the WTP. The economic value differences only occur between the higher income and lower income visitor groups; and this result is supported by the HCL model. As to the site estimations, Site 2 (Xinghai Bay) and Site 4 (Xinghai Beach) both exhibit the highest WTP for the attribute of water quality (WaterQ_ high) among all of the attributes analyzed, which means that the visitors have a higher desire for water quality improvement from the current condition. In terms of the recreation activities, visitors who choose to participate in swimming and boating show the highest desire for the water quality improvement (59.83 yuan), and the next is “beach walk” (46.32 yuan). All of the different types of tourists have a strong tolerance for crowding, since the lowest WTP is only -0.2 yuan. The improvement of the garbage situation has a significant influence on the WTP of tourists who prefer to sunbathe (39.1 yuan). However, according to the last row in Table 4, garbage amount (Garbage_med, 4-8 pieces in the view) generates the highest mean WTP among all of the samples. Overall, this implies that the visitors have a strong preference for a reduction in the amount of garbage, especially at Site 2 (Xinghai Bay).