Survey and choice experiment design

The data for this study was obtained from a survey conducted by the Everglades Foundation (a non-profit private foundation) in 2017 for the purpose of gathering information for its policy and communications strategy. The survey was administered by Qualtrics, a private market research firm, to a panel of adult respondents who volunteered in exchange for incentives (e.g. cash, gift cards, airline miles). All participants in panels used by Qualtrics must consent to being a part of the panel by submitting a form requesting to participate in market research. Potential respondents were sent an email invitation to complete the survey and were given information regarding the purpose of the survey, its basic content, and the estimated time needed to complete it. Participants could choose to end their participation at any time. The Everglades Foundation is not a Federalwide Assurance Institution (FWA) under U.S. Department of Health and Humans Services regulations nor were any federal government funds used to conduct the survey. Thus, U.S. federal regulations requiring Institutional Review Board (IRB) approval for research involving human subjects did not apply to administration of the survey (see 45 C.F.R. § 46, 2016). The survey and its administration adhered to the ethical and other standards set forth in the ICC/ESOMAR International Code on Market, Opinion and Social Research and Data Analytics.

The survey was comprised of five parts: (1) an introduction that explained the purpose of the survey and asked respondents to provide their sociodemographic characteristics (e.g., age, gender, race/ethnicity, and income); (2) a brief description of the Everglades, how it has changed from its original state, and current conservation efforts embodied in CERP; (3) a description of the ecological and cost attributes included in the choice experiment; (4) the SPCE questions and follow-up questions to test for biased or strategic responses; and (5) questions related to respondents’ political ideology and environmental attitudes. Environmental attitudes were measured using the New Ecological Paradigm scale [21,22] and other questions specific to Everglades restoration. Respondents were presented with a series of six choice scenarios with each choice scenario containing two restoration options (Choice A and Choice B) with differing levels of improvement in ecological attributes and an associated annual cost, and an opt-out option labeled Choice C with no ecological restoration and zero cost, which helps mitigate potential bias [23]. After making their choices, respondents were presented with a series of questions asking whether they thought cost should be a factor in ecological restoration of the Everglades, if they found the choice experiment questions difficult to answer, if they were concerned about the government’s ability to manage ecological restoration programs, and whether they thought it was fair for them to have to pay for restoration of the Everglades.

A set of potential attributes for the choice experiment were chosen based on a comprehensive literature review [13,15,17–20,24–36] and informal interviews with several hydrologists and ecologists familiar with the Everglades ecosystem who are engaged in research and/or management regarding this restoration effort. Attributes were evaluated based on their connection to CERP approved indicators, whether there were existing research or models published in the peer-reviewed literature that could be used to quantify how the attributes would respond to various restoration scenarios, and whether they could be easily understood by Florida residents. The attributes included in the survey were wading bird populations in Everglades National Park, the American alligator population in Everglades National Park, the population of the endangered snail kite in the greater Everglades Ecosystem (see Fig 1), the population of spotted seatrout in Florida Bay, and the amount of polluted discharges from Lake Okeechobee to the Caloosahatchee and St. Lucie rivers during wet periods.

Increased alligator populations are considered to be a system-wide performance metric for restoration of the hydrology in critical areas of the Everglades ecosystem [36]. Habitat suitability models for the American alligator—Alligator mississippiensis have been developed over the past several decades [20,26,28,29,31]. Wading birds are considered a performance metric for the resumption of the flow of fresh water through the southern Everglades [33]. A species distribution, foraging and reproduction model has been developed for wading birds including Great egrets (Ardea allea), white ibises (Eudocimus albus), and wood storks (Mycteria Americana) [17]. Snail kites rely almost exclusively on one prey item, the apple snail (Pomacea paludosa) [25]. Everglade snail kite populations are considered to be an important indicator of water quality in wetland ecosystems throughout the Greater Everglades [32]. A spatially explicit apple snail population model [18] has been developed that can be used to predict how restoration will impact Everglade snail kite populations. Florida Bay supports recreational fisheries for various species including spotted seatrout (Cynoscion nebulous) [27]. Spotted seatrout populations are considered a key indicator for the restoration of freshwater sheet flow into Florida Bay that are necessary for the alleviation of hypersaline conditions [35]. Spotted seatrout populations’ responses to salinity and water temperature can be predicted from an existing habitat suitability model [19]. Finally, during especially wet periods, water is discharged from Lake Okeechobee to the St. Lucie River (to the east) and the Caloosahatchee River (to the west). These polluted discharges lead to periodic toxic algae blooms, seagrass and oyster die-offs, and other detrimental ecological impacts [24,30]. Improvement in the ecological conditions of the St. Lucie and Caloosahatchee rivers and estuaries are considered a measure of successful water management in Lake Okeechobee as part of Everglades restoration [34]. The descriptions of these attributes that were presented to survey respondents are shown Table 1. Levels for the ecological attributes encompass realistic expected ecological responses to restoration as determined by a literature review and discussions with experts familiar with Everglades restoration and ecology. Table 1 also describes the levels of each of the attributes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Description of attributes and their levels presented to respondents in the choice experiment section of the survey.

https://doi.org/10.1371/journal.pone.0234051.t001

The payment vehicle used for the cost attribute was a tax on utilities. This vehicle was used because nearly all Florida households would have to pay a utility tax regardless of whether they owned or rented their place of residence and because this payment vehicle had been successfully used in previous surveys of Florida residents regarding Everglades restoration [13,15]. The payment levels were based on past surveys of people’s WTP for ecological benefits in general and Everglades restoration specifically, and were set at levels that would generate sufficient revenues to cover the expected costs of restoration. [15,37,38].

Owing to the number of attributes and attribute levels, a full factorial design that would include all possible combinations of attribute levels was not feasible. Alternatively, several design criteria can be used to generate an experimental design able to estimate a large number of parameters efficiently while keeping the survey short enough that respondents would be likely to complete it. These designs vary mainly in terms of orthogonality and efficiency. The software SAS [39] was used to create a D-efficient experimental design following the method outlined by Mitchell [40]. The design used in the survey consisted of eight blocks with six choice scenarios each for a total of 48 choice scenarios. Respondents were randomly assigned to one of the eight blocks. Fig 2 depicts an example choice scenario.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Example choice faced by survey respondent.

https://doi.org/10.1371/journal.pone.0234051.g002

The survey was iteratively tested and refined based on feedback from experts in Everglades ecology and choice experiments, then pretested on a sample of 100 representative Florida residents. The survey was fully implemented (n = 2,000) as a web-based survey to a balanced panel of Florida residents in December 2017. No substantive changes were made to the survey following the pretest, resulting in a usable sample size of 2,100.

Econometric specification and analysis

Discrete choice experiments are grounded in the random utility model (RUM) [41–44], if alternative j is chosen out of a set of alternatives i = 1, …., j, …J by respondent n in a particular choice situation t, then their indirect utility function can be represented as: (1) where X_njt is a vector of attributes associated with alternative j and β is a vector of coefficients associated with the attributes of alternative j and characteristics of respondent n. The term μ_njt represents the random (or unobservable) component of the utility of alternative j for respondent n.

Based on the random (from perspective of the researcher) component of utility, the probability of alternative j being chosen over another alternative s in choice situation t can be represented as a probability function where the probability of respondent n choosing option j assuming j ≠ s can be represented as Eq 2 [41,44]: (2) If the error terms in the above equations are independently and identically distributed and have a type I extreme value distribution, Eq (2) can be estimated using a conditional logit function [26,41]: (3)

The dependent variable in Eq (3) is the probability that a respondent will choose a specific alternative in a choice set. The conditional logit function has been widely used to model choice behavior in discrete choice experiments [41,43–45]. However, the conditional logit model makes three limiting assumptions. It assumes homogeneous preferences across individuals, independence of irrelevant alternatives, and no correlation of unobserved factors over time [46]. These assumptions are often violated in practice; thus researchers have increasingly turned to random parameters or mixed logit models to estimate WTP for improvements in environmental quality. The random parameter logit model relaxes all three of the above assumptions [41,43] and can be represented as follows: (4) where β_n represents a vector of coefficients associated with the attributes of j and characteristics of respondent n. In the random parameter logit model at least some of the attribute coefficients are assumed to be random and thus vary across respondents [43,47].

As the dependent variable in Eqs (3) and (4) is the probability that a respondent will choose a specific alternative in a choice set, the estimated coefficients of the independent variables estimate how that particular variable will influence the probability that an alternative is chosen. Because cost was one of the attributes, the WTP for the other attributes can be calculated by taking the negative of the ratio of the estimated coefficient of the attribute of interest, β_k, and the cost attribute as shown in Eq (5) [43]: (5)

When using the random parameter logit model to account for preference heterogeneity, a priori assumptions must be made regarding the distribution of the coefficients. Usually a normal or log-normal distribution is assumed. Thus, the calculation of WTP involves the ratio of two distributions where the denominator (the coefficient of the cost attribute, which is often assumed to be lognormally distributed) can typically take on values close to zero, resulting in the derived distribution of WTP being exceedingly large and unrealistic at the upper end of the distribution. This skewed distribution can bias the mean and variance of the WTP estimates [48–50]. An alternative to estimating WTP in preference space is to reparametrize the model so that the parameters in the model are marginal WTP for each attribute [48–50], i.e. the model is estimated in WTP space. Some empirical evidence suggest that models estimated in WTP space produce more reasonable distributions for WTP estimates [50]. To test the robustness of our estimates we estimated WTP using the conditional logit model, the random parameter logit model in preference space, and the random parameter logit model in WTP space. The alternative specific constant was labeled ‘optout’ and coded as 1 for the optout option and 0 for both of the restoration alternatives.

To better understand heterogeneity in preferences and the influence of respondent characteristics on respondent choices we used a latent class model where respondents were segregated into classes or segments based on their preference, attitudes, and sociodemographic characteristics. The important distinction of the latent class model relative to the random parameter model is that the vector of attribute coefficients for respondents in a class are the same (i.e., respondents within a class exhibit homogeneous preferences), but estimated coefficients vary between classes (i.e. different classes are heterogeneous in preferences) [15,41,51]. Following Pienaar et al and Milon and Scrogin [15,51], the latent class model can be represented as follows: (6) Where α is vector of parameters and γ is a vector of sociodemographic, political ideology, and environmental attitude variables that determine class membership. Eq (6) represents the probability of individual n choosing alternative j in choice situation t conditional on individual n belonging to class k. WTP is calculated the same way as in Eq (5). The latent class model with the number classes that best explained respondent heterogeneity while minimizing the risk of overfitting was determined by using the Consistent Akaike Information Criterion (CAIC) and the Bayesian Information Criterion (BIC) [51,52].

All regressions were performed using Stata 15.1 [53]. The built-in Stata clogit command was used to estimate the conditional logit model while the mixlogit and mixlogitwtp Stata commands developed by Arne Hole were used to estimate the random parameter logit models in preference space and WTP space, respectively [54,55]. Stata built-in commands and the user-written commands lclogit2 and lclogitml2 developed by Yoo [56] were used for the latent class analysis.

Respondent demographics

Overall, the sample was fairly representative of the population of Florida. A detailed comparison of the sample with Florida demographics is presented in S1 Table in the supplemantry material. The sample was more educated with about 52% of the sample with a bachelor’s degree or higher, compared to around 28% of the population. The sample also had a higher income with a median household income of $70,000 per year, compared to around $49,000 for Florida’s population. The survey did not ask respondents about their political party affiliation, but they were asked to identify themselves as liberal or conservative on a 7-point Likert scale, with one being the most conservative and seven being the most liberal. Even though it is recognized that the political party identification is imperfectly correlated with liberal/conservative ideology [57,58] if those self-identifying from one to three are classified as conservative, those identifying as a 4 as independent, and those identifying at 5–7 as liberal, then these classifications roughly aligned with the population identification of Republican, Independent, and Democrat respectively. Other demographic variables were similar between the sample and the population of Florida.

Logit model results

A description of the variables used in the conditional logit and the random parameter logit models are shown in Table 2. The results from the conditional logit model, the random parameter model in preference space, and the random parameter model in WTP space were broadly consistent with each other. For all three models, the estimated coefficients had the theoretically expected signs. The ecological attributes (increased populations of wading birds, American alligators, Everglades snail kites, spotted seatrout, and reduced polluted discharges from Lake Okeechobee) all had positive coefficients, meaning that their presence increased the probability of a restoration option being chosen, all else being equal. The estimated coefficient for the cost attribute was negative, meaning that all else being equal, the higher the cost the less likely it would be that a respondent would choose a restoration option. The estimated coefficient for the optout variable was negative, indicating that, overall, respondents preferred restoration of the Everglades over the status quo.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Description of variables used in regression models.

https://doi.org/10.1371/journal.pone.0234051.t002

The results for the random parameter logit models in both preference space and WTP space are shown in Table 3. Because all models produced similar results, the results from theconditional logit model are presented in S2 Table in the supplementary material. With the exception of a 50% increase in the American alligator population, coefficients estimated in preference space were significant at the 5% level. For the model estimated in WTP space, the coefficients for all attributes were significant at the 1% level. Both random parameter logit models indicated that respondents’ preferences for ecological restoration were heterogeneous, although the models were not entirely consistent in terms of which standard deviation coefficients were statistically significant. The magnitude of WTP estimates was similar across the conditional logit and random parameter logit models, with respondents showing the greatest WTP for reduced discharges of polluted water from Lake Okeechobee and the least WTP for increased American alligator populations in Everglades National Park. Although the WTP estimates were similar, empirical evidence suggests that estimating the random parameter logit model in WTP space instead of preference space produces more reasonable willingness to pay estimates [48,50]. That seems likely the case here as the WTP estimates from the random parameter models estimated in WTP space were all significant at the 1% level.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Random parameter logit results in preference space and WTP space.

https://doi.org/10.1371/journal.pone.0234051.t003

Willingness to pay for an increase in the American alligator and spotted seatrout populations remained approximately the same as the population level increased, with mean WTP at each population increase being well within the 95% confidence interval of the other levels. Respondents were willing to pay for a relatively small increase in these species populations but were relatively indifferent to further increases beyond an initial population increase of 10 percent. Respondents’ WTP for increases in the populations of wading birds and the Everglade snail kite increased when the expected increase in population size transitioned from 10% to 50%, but further population increases did not result in much change in WTP. Respondents’ WTP for reduced discharges of polluted water from Lake Okeechobee to the St. Lucie and Caloosahatchee rivers increased in a roughly linear fashion as reductions in polluted discharges increased.

Latent class model results

To better understand the nature of the preference heterogeneity among survey respondents, a latent class model specification was used where respondents were divided into distinct classes based on their responses to the New Ecological Paradigm (NEP) scale [22], other attitudinal questions, and respondent socioeconomic characteristics (see Table 4). The NEP scale has been widely used to measure environmental attitudes of survey respondents [59]. The scale consists of 15 statements where the respondent is asked to indicate their level of agreement or disagreement for each statement on a 5-point Likert scale (see Table 5). Eight of the 15 statements express sentiments that endorse an eco-centric or ecological world view, while the other 7 statements express sentiments that align with a more anthropocentric or “Dominant Social Paradigm” world view [22]. Previous research has shown that responses to the 15 NEP questions measure a latent attitude or worldview concerning how humans relate to the natural world [21,22,59].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. The sociodemographic and attitudinal questions used in the latent class model.

https://doi.org/10.1371/journal.pone.0234051.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Questions composing the new ecological paradigm scale.

Agreement with unshaded questions indicate a “eco-centric” view, while agreement with shaded questions indicate a “anthropocentric” world view.

https://doi.org/10.1371/journal.pone.0234051.t005

The responses to the NEP scale statements can be aggregated into a single score ranging from 15 to 75 to indicate variation in respondents’ underlying attitudes, with a higher score indicating a more eco-centric attitude. Fig 3 shows a histogram of aggregate NEP scores from the survey respondents displaying an approximately normal distribution with a slight skew toward an eco-centric world view. This result is consistent with other studies using the NEP scale [22,60–62]. The interitem correlations, which measure if individual questions on a questionnaire, are correlated) and Cronbach’s alpha, which measure the internal reliability of survey questions, are shown in Table 6. The average interitem correlation was 0.329 on the test scale which is within the ideal range of 0.15 to 0.50 and the average Cronbach’s alpha was 0.820 which shows a good level of internal reliability. The Cronbach’s alpha score on each individual item indicates how much the average Cronbach’s alpha would decrease if that item was removed. As can be seen in Table 6, removing any of the items would decrease the average, indicating that no item should be removed from the scale to improve reliability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Histogram of NEP scores.

Statements expressing an “anthropocentric” view were reverse coded. Scores range from 15 to 75 with higher scores indicating a more “eco-centric” attitude toward nature.

https://doi.org/10.1371/journal.pone.0234051.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Interitem correlations and Cronbach’s alpha for NEP questions.

https://doi.org/10.1371/journal.pone.0234051.t006

Results from the latent class model confirm that there is preference heterogeneity among the survey respondents. According to the Consistent Akaike Information Criterion (CAIC), a model with four classes provides the best fit, and according to the Bayesian Information Criterion (BIC), a model with five classes provides the best fit while minimizing the risk of overfitting. To use the simplest model and ensure that the model did not overfit the data, we used a four-class model. Results of the latent class model are presented in Table 7 and Table 8 presents the willingness to pay estimates calculated using Eq 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Results from the latent class model.

https://doi.org/10.1371/journal.pone.0234051.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Willingness to pay estimates for each class from the latent class model.

https://doi.org/10.1371/journal.pone.0234051.t008

As can be seen in Table 7, class 3 is the largest with 38.3% of the respondents while class 1 contains the least with 15.6%. Classes 2 and 4 contain 22.5% and 23.5% respectively. Class 4 was set as the base for interpretation of the coefficients of the sociodemographic and attitudinal variables used to determine class membership. All of these variables, except one, were significant at the 5% level for at least one of the classes, meaning that at least one class differed from class 4 on these variables. The only exception was the variable other_race for which none of the first three classes differed significantly from class four. The coefficient for the optout choice was significant for all of the classes. Classes 2, 3, and 4 had negative coefficients, indicating that respondents in these classes preferred restoration over the status quo. Class 1 had a positive coefficient for the optout, meaning the opposite for these respondents (i.e., they do not value Everglades restoration). The coefficient for the cost attribute was significant, at least at the 10% level, for classes 1, 3, and 4 indicating the higher the cost of a restoration option, the less likely the respondents in these classes would choose it. The coefficient for the cost attribute was small and not statistically significant for class 2 indicating that respondents in this class displayed attribute non-attendance for this attribute, meaning that cost was not a factor in choosing among alternatives. This resulted in a very small, non-significant coefficient for cost for respondents in this class. Since the coefficient for the cost attribute is the denominator in calculating WTP (see Eq 5), the willingness to pay estimates for the restoration attributes for these respondents were also not significant. [63–65][65]

To facilitate discussion, the attitude and demographic characteristics of each of the four classes are shown in Table 9. Class 1 skewed whiter (relative to classes 3 and 4) and more male with fewer respondents with a college degree. This class also displayed a more anthropocentric attitude based on their responses to the NEP statements and were less likely to agree that cost should not be a factor in restoration of the Everglades and more likely to agree that they should not have to pay additional fees or taxes for Everglades restoration relative to the other three classes. As indicated previously, this class had a very small or no willingness to pay for the restoration attributes and indicated a strong preference for the status quo instead of paying for restoration. Class 2 also skewed whiter and more male. However, this class had a larger number of respondents with a college degree and on average had higher income compared to the other classes. This class displayed more eco-centric views according to their responses to the NEP statements, and they self-identified as more liberal relative to other segments in terms of political ideology. However, the willingness to pay for restoration attributes for this class was not statistically significant. Classes 3 and 4 skewed more female and indicated willingness to pay for most of the restoration attributes. These classes also had higher (or more ecocentric) scores on the NEP scale relative to class 1. The difference between these two classes is that class 3 skewed more non-white and younger while class 4 had a higher willingness to pay for reduced discharges relative to class 3. Somewhat surprisingly class 3 also self-identified as more conservative than other classes in terms of political ideology.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Attitude and demographic characteristics of the four classes resulting from the latent class model.

Refer to Table 6 for descriptions of the variables and their coding.

https://doi.org/10.1371/journal.pone.0234051.t009