Background

Freshwater resources are an integral part of the cultural heritage, economic development, and national character of New Zealand (NZ) (Ambrey et al., Reference Ambrey, Fleming and Manning2017; Awatere et al., Reference Awatere, Robb, Taura, Reihana, Harnsworth, Te Maru and Watene-Rawiri2017). The indigenous Māori culture is important in NZ, and from a Māori world view, the separation of Ranginui (sky father) and Papatūānuku (earth mother) produced freshwater, emphasizing both the importance and connection of freshwater to the Māori people and NZ population more broadly (MFE & Stats NZ, 2020). Given this importance, there are several existing water quality valuation studies, but many are unpublished, in the gray literature, or from government (Miller, Reference Miller2014; Phillips, Reference Phillips2014; Tait et al., Reference Tait, Miller, Rutherford and Abell2016) or consultant reports that do not yield original estimates (Marsh and Mkwara, Reference Marsh and Mkwara2013). Other studies focus on only one region of NZ (Tait et al., Reference Tait, Baskaran, Cullen and Bicknell2011; Marsh and Phillips, Reference Marsh and Phillips2015). Transferring the results of these case studies to policies and populations in other areas or at the national level is questionable and could result in large errors (Smith and Pattanayak, Reference Smith and Pattanayak2002).

The NZ-based SP literature is similar to the international literature (Van Houtven et al., Reference Van Houtven, Powers and Pattanayak2007) in that it has examined several different water quality indicators. Some of these indicators are related to agricultural practices, such as riparian buffer restoration (Cullen et al., Reference Cullen, Hughey and Kerr2006) or nutrient leaching (Baskaran et al., Reference Baskaran, Cullen and Colombo2009; Takatsuka et al., Reference Takatsuka, Cullen, Wilson and Wratten2009). Tait et al. (Reference Tait, Baskaran, Cullen and Bicknell2011) value the ecological condition of waterbodies using poor, fair, and good quality categories. These categories are described by the type of weeds present, percent algae cover, and the types of insects and fish species present. Swimming suitability indicators of waterbodies have also been used (Marsh, Reference Marsh2012; Miller, Reference Miller2014; Miller et al., Reference Miller, Tait and Saunders2015) to represent recreation and health impacts. Marsh and Phillips (Reference Marsh and Phillips2015) used several different indicators of water quality alongside a qualitative swimming suitability measure, including ecological health, salmon and trout condition, and tributary water quality, which were presented qualitatively as good, satisfactory, not satisfactory, or poor. Translating many of these indicators and qualitative categories to marginal changes, as often predicted from policy projections, is difficult and generally inappropriate. Tait et al. (Reference Tait, Miller, Rutherford and Abell2017) use qualitative indicators like poor, moderate, and good for water clarity and ecological quality. However, they directly link each of their attributes to objective ranges in the underlying water quality parameters. For instance, poor ecological quality is defined as a Macroinvertebrate Community Index score less than 80, and poor clarity is defined as visibility of less than 1.1 m. It is not clear, however, if those quantified ranges were presented to respondents.

Many of these water quality indicators are used in other studies internationally (US EPA, 2015; Johnston and Bauer, Reference Johnston and Bauer2020). The US EPA commonly uses a Water Quality Index (WQI) for valuation work on many rules, which combines several other parameters, including nutrients, pH, temperature, clarity, dissolved oxygen, and sediment (Walsh and Wheeler, Reference Walsh and Wheeler2013). Meta-analysis benefit transfer is used for EPA valuation, which translates studies that use other water quality indicators into the WQI (Johnston et al., Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa, Tourangeau and Vossler2017). The European Framework Directive aims to improve waterbodies to “good” status, as defined by several underlying water quality variables. SP studies there have focused on both individual parameters like nutrients and the bad/poor/good status of waterbodies (Ferrini et al., Reference Ferrini, Schaafsma and Bateman2014; Anciaes, Reference Anciaes2022).

Johnston et al. (Reference Johnston, Schultz, Segerson, Besedin and Ramachandran2012) provide guidelines for including ecological content in SP surveys. They note that less structured treatment of attributes can cause problems with subsequent welfare estimation. Respondents’ internal conceptualization of the commodity may be different from that presented or intended by the researchers. This can be a complicated balancing exercise with water quality because the commodity itself has multiple dimensions that can be difficult to communicate to survey participants. For instance, Milon and Scrogin (Reference Milon and Scrogin2006) explored values associated with wetland restoration by presenting respondents with either functional attributes like water levels or structural attributes like species abundance, with the latter yielding significantly different (higher) WTP. In their paper on contemporary SP guidance, Johnston et al. (Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa, Tourangeau and Vossler2017) recommend that “the change being valued be based on how respondents tend to perceive the good.”

The size of the change presented to respondents creates another issue with applying previous SP estimates for benefit transfer. Miller et al. (Reference Miller, Tait and Saunders2015), for example, have respondents compare policies that result in 0%, 20%, or 40% improvements in the percent of sites suitable for swimming. There are few plausible policies that could improve water quality (or reduce nutrient inputs) by that large of a change. Nonetheless, such large changes, on the order of 20% to 50%, are often applied in SP surveys (e.g., Baskaran et al. [Reference Baskaran, Cullen and Colombo2009]). A meta-analysis of 140 observations from 51 SP studies of water quality (USEPA, 2015), where quality was represented on a scale of 0–100 with 100 representing pristine waterbodies, found that less than 10% of the observations used water quality indices measures under 10 (Figure 1).Footnote ²

Figure 1. Distribution of Water Quality Index (WQI) changes in 51 stated preference surveys (USEPA, 2015).

Survey design and implementation

We considered all the aforementioned gaps when designing the SP survey instrument used in this study, including: identifying water quality measures and changes that matter to the general public and that could be accurately understood by respondents, are realistic, and that could be directly linked to objective policy-induced changes. The ultimate commodity being valued in the SP survey is improvements in the quality of rivers and streams in the regional council where a respondent resides. The survey was implemented in 2018 and 2019. There are two versions of the survey, one for the North Island and one for the South Island of New Zealand. The surveys are identical except for bar graphs illustrating current attribute levels for each region on that island.

The survey instrument was developed and refined using focus groups and cognitive interviews.Footnote ³ Six focus groups were conducted in total during May and June 2018 at three different locations: two focus groups each at two urban locations (Auckland and Wellington) and two in a rural area (Hawke’s Bay). Input from the focus groups was used to refine the survey text and questions, identify relevant water quality attributes, and improve communication and presentation. To further refine the survey instrument, 10 cognitive interviews were conducted. Eight of the cognitive interviews were in Wellington and two were held in the rural Wairarapa area.Footnote ⁴

Depending on where a respondent lives, the survey begins with a map of the North or South Island that includes the regions and the major rivers on that island. To emphasize consequentiality (Carson and Groves, Reference Carson and Groves2007) and credibility of the survey, the instructions remind respondents’ that their “… answers will help inform policy makers” and that the baseline data are provided by the MFE and regional council governments. Respondents are then asked questions on recreational use and visitation to rivers and streams in their region, followed by introductory text defining each water quality attribute. Respondents are then shown figures depicting the current baseline levels in their regional council area for each attribute, as well as for the other regions of their island. Questions about awareness of the attributes are also presented. See the full survey example in Appendix D for details.

The policy scenarios, provision mechanism, and payment vehicle for public programs to improve water quality in a respondent’s region are then introduced. To minimize hypothetical bias and enforce consequentiality (Johnston et al., Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa, Tourangeau and Vossler2017; Vossler and Zawojska, Reference Vossler and Zawojska2020), respondents are reminded to act as though their household is actually facing the costs presented and that their responses could influence future policies and programs, as well as costs to their household. A generic regional policy is described as the provision mechanism.

The payment vehicle is specified as a permanent increase in a household’s general cost of living. More specifically, respondents are told that their monthly cost of living would change due to increases in home maintenance costs, utility bills, rent, and/or the price of food and other goods.

The survey then presents respondents an example choice question, followed by three separate discrete choice questions. Each choice scenario includes a status quo option and two policy alternatives. In the status quo option, water quality attributes remain at their current levels, while in the two policy options there are improvements in one or more of the water quality attributes, as well as an associated permanent increase in monthly living costs. Respondents are instructed to consider each choice question independently.

The final survey instrument includes three water quality parameters: water clarity, nutrients, and E. coli. These parameters appear in several upcoming NZ freshwater policies and are well known to the public (MFE, 2020). Statistics NZ includes these parameters in their list of central water quality-tracking indicators (see https://www.stats.govt.nz/indicators/). The representation of each parameter is also chosen to match policy levers and is informed by input from the focus groups and cognitive interviews. Water clarity is expressed as the average visibility for rivers and streams in a respondent’s region and is measured as Secchi disk depth. Nutrients are measured as the percent of rivers and streams in the region that have nutrient levels considered acceptable for aquatic life by official nutrient limits.Footnote ⁵ Similarly, E. coli is measured as the percent of rivers and streams in the region where concentrations are low enough to be considered suitable for swimming, wading, and fishing.Footnote ⁶ Each of these attributes was introduced with several sentences of explanation. For instance, with nutrients, respondents were told that although nitrogen and phosphorous are naturally occurring, too much can lead to excessive algal growth that harms underwater habitat (the full descriptions appear in the appendix). The text in the explanations was refined through the focus groups and cognitive interviews.

While the three water quality attributes described are correlated and the ecological end points that individuals directly care about may relate to more than one attribute in some cases (Tait et al., Reference Tait, Baskaran, Cullen and Bicknell2011), distinctions are drawn in the survey as to what each attribute primarily reflects. Excessive levels of nutrients are described as adversely impacting aquatic ecosystems, although reduced esthetics are also mentioned. Water clarity is described as how “murky or cloudy” the water visually appears. E. coli is described in terms of how it affects the health of people who swim, wade, or fish in the water. Based on the past literature and our own focus group and cognitive interview findings, these categories reflect what the general public find to be the most relevant end points related to surface water quality.

The size of the water quality changes presented to respondents are based on the magnitude of changes in national targets from the National Policy Statement for Freshwater Management (NPSFM; (MFE, 2020), p. 64). For example, Table 1 shows the national targets for improvements in primary contact suitability across different waterbody categories, with red being the lowest quality rivers and blue being the highest quality rivers (the full NZ Government figure this is drawn from appears in Appendix C). For example, between 2017 and 2030, the goal is to have a five percentage point reduction in rivers in the worst category (red) and a three percentage point increase in rivers classified in the highest category (blue). These changes are smaller than the scenarios presented in many previous SP surveys (US EPA, 2015).

Table 1. National targets for improvements in primary contact waterbodies

Each attribute and the posited changes in the attribute levels in the survey are presented in Table 2. A Bayesian efficient design was developed for the choice questions using Ngene software (ChoiceMetrics, 2018). Although there is not much information on the priors for each coefficient, the sign of each parameter was informed by past NZ-based literature (Marsh et al., Reference Marsh, Mkwara and Scarpa2011; Tait et al., Reference Tait, Baskaran, Cullen and Bicknell2011, Reference Tait, Miller, Rutherford and Abell2017; Marsh and Phillips, Reference Marsh and Phillips2015). One advantage of Bayesian efficient designs is that they are more robust than other designs to mis-specification of the priors (ChoiceMetrics, 2018).

Table 2. Changes in survey attribute levels

The SP survey module concludes with a series of questions to gauge the respondent’s perceived consequentiality of their responses and flag individuals potentially exhibiting protest and warm-glow behaviors. Such questions are used to screen the sample of respondents exhibiting potentially biasing behaviors and assess the robustness of our results. The broader survey includes socioeconomic questions to allow us to examine preference heterogeneity and possibly further tailor such heterogeneity when extrapolating benefit estimates to the broader population.

An example choice question from the survey appears in Figure 2.

Figure 2. Example choice question.

The survey was administered online by Horizons Research as a separate module in a broader survey on environmental preferences in NZ that is implemented every few years by Lincoln University (Hughey et al., Reference Hughey, Kerr and Cullen2019). Horizon Research maintains an internet panel of approximately 7000 people. The survey was open from March to April 2019, and 2007 respondents participated. When compared to population data from the Census, our survey sample overrepresents individuals 60–69 years of age, those with a tertiary education, and urban populations, and underrepresents individuals 18–19 years of age, people with only high school qualifications, and rural populations.Footnote ⁷

A key assumption when extrapolating survey results is that the estimates are representative of the general population. We attempt to bolster this assumption in two ways in order to provide estimates that can be more defensibly applied in benefit transfer exercises. First, we weight responses by regional council area population totals to better reflect the spatial distribution of the population across New Zealand, as well as any systematic differences in those populations and their preferences (see section “Data” for details). Second, in our more comprehensive models, we parametrically control for preference and income heterogeneity and then use the population data from the Census to predict more representative average WTP estimates (see sections “Results” and “Policy illustration” for details). Despite our best efforts, it is nonetheless important to keep this key assumption in mind when interpreting and extrapolating our results.

Methodology

We employ a random utility model (RUM) framework to analyze the data from this discrete choice experiment. In these models, utility is divided into a deterministic component and a random component, represented by v(.) and ϵ, respectively. The utility that household i receives from alternative j is

(1) $${u_{ij}} = v({{\bf{a}}_j},{I_i} - {C_j}) + {\varepsilon _{ij}}$$

This specification assumes that the first component of utility is a function of the group of attributes defining each alternative ${a_j}$ , along with numeraire consumption ( ${I_i} - {C_j}$ ), which is the difference between household income ${I_i}$ and the cost of the alternative ${C_j}$ .

In the empirical models, we also add a status quo constant ( $sq{c_i}$ ), which represents respondents’ preferences for or against the status quo option in general, irrespective of the attributes defining the alternative policy options. We allow for unobserved heterogeneity in the $sq{c_i}$ by estimating it as a normally distributed random parameter in a mixed logit framework. In doing so, we accommodate both respondents that have a bias toward or against the status quo (Moore et al., Reference Moore, Guignet, Dockins, Maguire and Simon2018).

We assume a linear specification for v(.). RUMs are often estimated as conditional or mixed logit specifications (Greene, Reference Greene2000; Haab and McConnell, Reference Haab and McConnell2002). The conditional probability that household i would choose alternative j appears in equation (2):Footnote ⁸

(2) $${P_i}(j|{a_n},{C_n}) = {{\exp (sq{c_i}{D_j} + {\bf{\beta }}{a_j} + \delta {C_j})} \over {\sum\nolimits_n {\exp (sq{c_i}{D_n} + {\bf{\beta }}{a_n} + \delta {C_n})} }}$$

In this formulation, n refers to alternative options in a given choice occasion, D is an indicator variable denoting the status quo alternative, $\beta $ is a vector of coefficients to be estimated, and δ is the coefficient on the cost attribute. δ can be interpreted as the negative of the marginal utility of income. We explore individual-level preference heterogeneity in two ways. First, we include several interaction terms between the main choice attributes and observed household characteristics, including household-specific socioeconomic variables like income and household size, recreational user-related variables, and baseline regional water quality. Second, we explore possible unobserved preference heterogeneity by allowing $\beta $ to vary as a random coefficient across respondents with each element of $\beta $ following an independent normal distribution (Mariel and Artabe, Reference Mariel and Artabe2020).Footnote ⁹ The cost parameter δ is held fixed to ensure MWTP has defined moments (Layton and Brown, Reference Layton and Brown2000, Revelt and Train, Reference Revelt and Train2001, Daly et al., Reference Daly, Hess and Train2012). Alternate approaches estimate models in WTP space, in which the distribution of welfare is modeled (Train and Weeks, Reference Train and Weeks2005, Scarpa et al., Reference Scarpa, Thiene and Train2008). However, the latter approach is not always found to fit the data as well, and there are computational challenges with estimation (Johnston et al., Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa, Tourangeau and Vossler2017).Footnote ¹⁰ The mixed logit models and subsequent calculations are estimated using Stata statistical software (StataCorp, 2021).

Welfare measures can be inferred from the estimated parameters. For example, under the linear specification in equation (2), the vector of household marginal willingness to pay (MWTP) estimates can be calculated as:

(3) $$MWTP = - {\beta \over \delta }$$

Given a projected policy change in the attribute levels from the baseline of ${a^0}$ to ${a^1}$ , and based on our assumed linear functional form, we can calculate the nonmarginal welfare change for a household as:

(4) $$WTP = MWTP \times \left( {{a^1} - {a^0}} \right)$$

Notice that we exclude the $sq{c_i}$ estimates from our welfare calculations. This status quo term captures a respondent’s tendency to favor or disfavor the status quo option irrespective of the improvements and costs defining the alternative policy options. The status quo term could therefore be capturing alternative, omitted variable biases that would otherwise confound the welfare parameter estimates of interest (Johnston et al., Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa, Tourangeau and Vossler2017; Moore et al., Reference Moore, Guignet, Dockins, Maguire and Simon2018). For example, $sq{c_i}$ could be capturing the “warm-glow” associated with choosing environmental action. That choice, in and of itself, however, is not directly related to the proposed policy change or attribute levels. Alternatively, this term could be partially capturing legitimate preferences for or against a policy and in that sense could be included it in welfare calculations. We speculate that $sq{c_i}$ likely captures both effects in most applications.

Nonetheless, the appropriateness of including $sq{c_i}$ in welfare calculations for benefit analysis remains unclear for two reasons. First, what respondents perceive to result from a policy option outside of the specified attribute changes is unobserved to us as the researchers. Respondents could be considering changes in end points that are in no way related to policies of interest. Second, our primary objective is to provide estimates for future benefit transfer. As is the intent here, government agencies often transfer primary study estimates to numerous, iteratively implemented policies (Petrolia et al., Reference Petrolia, Guignet, Whitehead, Kent, Caulder and Amon2021). This is due to the high costs of conducting appropriate original studies and a desire to streamline benefits analyses. As is the case in most of the SP literature, respondents in our study are asked to independently evaluate choice occasions where only a single policy, at most, would be implemented. Even if $sq{c_i}$ captured only legitimate preferences against the status quo, or for a policy – an assumption we do not believe would ever hold in reality – any WTP calculations that include $sq{c_i}$ could not be validly transferred to subsequent policy changes.Footnote ¹¹ To be conservative and ensure that the welfare calculations are as unbiased as possible, we exclude $sq{c_i}$ from the welfare calculations.

Data

Of the 2007 respondents that took the survey, 1736 completed all three choice questions (86%), 26 completed two (1.3%), and 7 respondents completed only one (<1%). The remaining 238 respondents (12%) did not respond to the choice questions and are excluded from the analysis. Among the 1769 respondents that answered at least one choice question, 73% are from the North Island (especially Auckland, Bay of Plenty, Waikato, and Wellington), 24% of the respondents are from the South Island (and in particular, Canterbury), and 3% did not provide their region (Table 3).

Table 3. Sample size by regional council in the North and South Islands

Note: Among the n = 1769 respondents, 59 did not provide information on the region where they live and are excluded from the above table.

To reduce the potential influence of biasing behaviors sometimes associated with SP methods, we screen the sample based on a combination of responses to the choice and debriefing questions. Based on the criteria below, we identify and flag respondents as potentially exhibiting the following behaviors:

• Consideration of other waters omitted from the choice experiment: Respondents who disagreed with the statement that they only considered rivers and streams in their region.

• Hypothetical bias due to warm-glow: Respondents who always chose the highest cost option in each choice question they were presented, and who agreed with the statement that it is important to improve water quality no matter how high the costs.

• Treated responses as nonconsequential: Respondents who disagreed with the statement that they made their choices as if the presented water quality improvements and increased costs would actually be experienced.

• Protest response: Respondents who always chose the status quo option and who agreed with one of the following statements: (i) that they value water quality improvements but their household should not have to pay for it, or (ii) that they are against more regulations and government spending.

Table 4 shows how the sample size changes as we screen out respondents exhibiting responses that one would expect to bias MWTP upward (going from left to right), and that could bias MWTP downward (going from top to bottom). The diagonal displays the sample sizes as we treat potential upward and downward biases symmetrically (Banzhaf et al., Reference Banzhaf, Burtraw, Evans and Krupnick2006; Moore et al., Reference Moore, Guignet, Dockins, Maguire and Simon2018). The upper left-corner shows the full sample size of 1769 respondents who answered at least one choice question and the bottom-right corner shows that 1364 respondents remain after fully screening out those who were flagged as potentially exhibiting biasing behaviors.

Table 4. Number of respondents in sample under alternative screening criteria

When estimating the regression models, observations are weighted to account for differences in sampling intensity, response rates, and sample screening across regions. We weight the observations in our regression models to ensure that the sample appropriately represents the population across the regions, which in turn allows interpretation of the estimates as national averages. The weight assigned to each respondent is the total population in their council region divided by the region-specific sample size after screening.

The survey also included several questions about respondents’ recreational activities in rivers and streams and respondents’ awareness of existing water quality levels in their region. Respondents were asked about activities they did at rivers and streams in their regional council area in the last 12 months and could choose multiple options from the following categories:

• Swimming or wading

• Fishing

• Boating, including sailing, and motor boating

• Water skiing, jet skiing, or kayaking

• Actively viewing nature (e.g., bird watching)

• Biking or walking on trails/paths alongside the water

• I didn’t visit rivers or streams in my regional council area in the last 12 months

The responses to these questions were aggregated into three user categories: contact users (including water skiing, jet skiing, or kayaking, swimming, or wading), non-contact users (including fishing, sailing, or motor boating), and passive users (those actively viewing nature, biking, or walking). Respondents can fall into more than one user category. After respondents were presented with baseline graphs and explanations of each water quality parameter, they were asked if they were aware of the characteristics or impacts of nutrients and E. coli, and whether clarity levels met their expectations. Table 5 summarizes the percent of respondents that fall into the user categories and percent of respondents who were aware of existing water quality levels.

Table 5. Types of users of rivers and streams and percent aware of existing water quality levels by regional council

There is some noticeable variation in how respondents use the rivers and streams in their regions (Table 5). For example, Nelson and Marlborough are areas known for their beaches and coastal amenities and so it is no surprise that a high proportion of respondents engage in water contact recreation in rivers and streams as well. Respondents in the West Coast Region also had very high participation in recreation, although the number of respondents there was small (see Table 3).

Policy illustration

To demonstrate how these values might be applied in a policy setting, we perform a benefit transfer on a simulated national water quality improvement that was previously modeled by the National Institute of Water and Atmospheric Research (NIWA) (Hicks et al., Reference Hicks, Greenwood, Clapcott, Davies-Colley, Dymond, Hughes, Shankar and Walter2016, Reference Hicks, Semadeni-Davies, Haddadchi, Shankar and Plew2019). Sediment was identified as a high-priority freshwater contaminant to manage. The National Policy Statement for Freshwater Management (NPSFM) did not previously have sediment as a target, so the MFE was interested in identifying the impact of proposed catchment sediment load limits (MFE, 2020). Catchment load limits could be achieved through land use conversions (such as converting erodible pasture into forestry) and other erosion best management practices aimed at reducing sediment from reaching waterbodies. Both in-stream sediment criteria and clarity criteria were formulated that would meet nationwide “bottom lines” in each of these four primary state variables.Footnote ¹⁵ We use the NIWA modeling data on clarity that project feasible improvements in clarity as a result of catchment load limits. The modeling identifies streams and rivers with a median clarity, that is, below the threshold, and simulates the potential improvement from the practices aimed to reduce catchment sediment loads.

The water quality improvements for each stream/river reach or segment s are weighted its length (Length _sr ), and then added together to get the reach-weighted average clarity change for each regional council area r; as in the following equation, where ${N_r}$ denotes the total number of river segments in region r:

(5) $$Mean\;Clarity\;Chang{e_r} = \mathop \sum \nolimits_{s = 1}^{{N_r}} \left( {{{Lengt{h_{sr}}} \over {\left( {\mathop \sum \nolimits_{s = 1}^{{N_r}} Lengt{h_{sr}}} \right)}}Clarity\;Chang{e_{sr}}} \right)$$

A summary of these average clarity improvements for each regional council area appears in Figure 4. Most of the regional councils see a small average change in clarity, of under 0.1 m, with the largest change in Waikato, at 0.154 m. These changes are proportionally smaller than the changes desired by the national policy statement, pictured in Appendix C, with some changes smaller than those presented in our choice experiment questions. This exercise further illustrates the difficulty in achieving long-term goals for water quality.

Figure 4. Projected mean regional council clarity improvements (m).

Although we have data on changes in clarity, we also need changes in nutrients and E. coli. This is a common problem with monetizing water quality policy: the need to convert between different parameters (Walsh and Wheeler, Reference Walsh and Wheeler2013; US EPA, 2015). It is likely that the policies used to improve sediment or clarity will also improve E. coli and nutrients. For instance, to achieve sediment load targets, Neverman et al. (Reference Neverman, Djanibekov, Soliman, Walsh, Spiekermann and Basher2019) simulate the impact of whole farm planning and afforestation, which will also improve E. coli and nutrient leaching to waterways.

To calculate the subsequent changes in E. coli and nutrients associated with the clarity improvements, we use data from NZ Statistics, who publish modeled segment-level data on several water quality parameters, resulting in almost 600,000 observations for each parameter.Footnote ¹⁶ For E. coli, total nitrogen, and total phosphorous, we use a regression to model the reduced-form relationship between each indicator (WQ) and clarity, as shown in equation (6). Several control variables are included in X: elevation and dummies for stream order and the dominant surrounding land cover. The regression also includes regional council fixed effects, and the E. coli regression includes dummies for the baseline “letter grade” of the stream (Appendix C). Note that to properly model the ecological relationship between these variables, a more in-depth approach should be used. However, for the purposes of this benefit transfer, these reduced-form regressions establish a reasonable relationship:

(6) $$\ln \left( {WQ} \right) = \beta \ln \left( {Clarity} \right) + \delta X + \gamma + \varepsilon $$

Regression results appear in Appendix E and exhibit significant negative relationships between the natural log of clarity and each indicator. The estimated relationship with E. coli is slightly lower than Davies-Colley et al. (Reference Davies-Colley, Valois and Milne2018); however, they use a simple correlation coefficient of −0.54. Our estimated coefficients are used to translate the change in clarity into the other indicators. Using the government thresholds referenced above, we can then determine which changes result in a segment moving from exceeding the threshold to not exceeding. For instance, with the E. coli government criteria, river and stream segments are assigned a letter grade from A to E, with D and E being unsafe for swimming.Footnote ¹⁷ If the forecast E. coli change bumps the waterbody from unsafe for swimming to safe, it is counted as no longer exceeding the unsafe threshold. The TP and TN results are combined into a nutrients indicator so that if either exceeds its threshold the waterbody is still counted as exceeding the acceptable limit for nutrients. The projected increase in rivers meeting the safe/acceptable E. coli and nutrient criteria appear in Figure 5. The values in Figures 4 and 5 highlight the difficulty in achieving meaningful water quality changes. Only 3–5 regional councils in each graph see water quality changes that are within the scope of the attributes presented in our survey (see Table 2).

Figure 5. Projected improvements in the percent of rivers meeting E. coli and nutrient criteria.

The average clarity changes are monetized using the results of our preferred Model (5), the regional council level averages of the relevant interacted variables for each model, and the number of households from Stats NZ.Footnote ¹⁸ Based on equation (4), the annual benefits for regional council area r in year t is calculated for each water quality attribute, using clarity as an example in equation (6), where the middle term is the mean clarity (or E. coli or nutrients) change from equation (5):

(7) $$Annual\;Benefit{s_{qrt}} = MWT{P_{qr}} \times Mean\;Chang{e_{qr}} \times H{H_{rt}}$$

where q denotes the corresponding water quality attributes (clarity, nutrients, or E. coli), and $H{H_{rt}}$ is the number of households in regional council area r in year t. The estimated average annual benefits at the household level from the change in each quality attribute appear in Table 8, with regional council-level benefits in Table 9. Waikato had the largest estimated clarity changes (Figure 5), which translate into the highest household-level benefits for two of three water quality attributes (Table 8). The household-level nutrient benefits for Waikato are notably higher than others due to the large MWTP for nutrients and the largest nutrients policy change. At the regional council level (Table 9), the largest total benefits accrue in Waikato, with approximately $46.6 million in benefits in our preferred model. Marlborough had the lowest annual benefits at $45 thousand.

Table 8. Annual household-level benefits from policy illustration (in NZD)

Table 9. Regional council-level annual benefits (in NZD)

The estimated national annual benefits of this policy change is approximately $115 million (Table 9). To illustrate the sensitivity of our overall estimates to model choices, Table 10 contains the total national benefits based on each model. The table shows the highest benefit estimates from the more parsimonious models (1) and (2). Our preferred model, the model that best fit the data based on BIC and AIC, yields total benefit estimates squarely in the middle of all our specifications.Footnote ¹⁹

Table 10. National annual benefits across models (in NZD)