1. Introduction
Agricultural nonpoint source (agNPS) pollution,Footnote 1 the primary cause of water quality deterioration, is a major concern for water-stressed countries (Bouwer, Reference Bouwer2000; Zhang et al., Reference Zhang, Luo and Zhang2024), including South Africa.Footnote 2 This pollution, which results from unsustainable agricultural practices, such as the intensive use of inorganic fertilizers and pesticides, is worsening due to the expansion and intensification of agriculture in response to population growth, increasing food demand and changes in dietary patterns (FAO, 2017). If left unchecked, this pollution poses severe consequences for water resources, including (1) degradation of water quality at the farm level; (2) serious eutrophication in water bodies, precluding water use for other key sectors in the economy, including industry, fisheries and recreation services (Halliday et al., Reference Halliday, Skeffington, Bowes, Gozzard, Newman, Loewenthal, Palmer-Felgate, Jarvie and Wade2014); (3) exacerbation of water scarcity that endangers water security goals and (4) salinization of water resources, leading to decreased agricultural yields, farm profits and irrigation efficiency, consequently jeopardizing agricultural production and food security. Reducing agNPS pollution is crucial, with numerous benefits for farmers and society as a whole. Its reduction would contribute to alleviating water scarcity and improving its quality, management and sustainability. This would enable farmers to continue to make meaningful contributions to economic growth and the welfare of society.
Both mandatory and voluntary measures are employed to control agNPS pollution. For mandatory measures, higher taxes on pollution-causing inputs, tax-subsidy schemes and direct regulations forcing farmers to adopt specific pollution reduction technologies (Feather and Cooper, Reference Feather and Cooper1995; Segerson, Reference Segerson1988; Xepapadeas, Reference Xepapadeas1997) are often used. However, due to the characteristicsFootnote 3 of agNPS pollution, mandatory measures are often ineffective and costly to implement (Han and Zhao, Reference Han and Zhao2010). In contrast, voluntary incentives, such as cost-sharing, incentive payments and technical assistance, are considered a preferred approach as they motivate farmers to adopt practices that improve water quality. Many countries have incorporated voluntary incentives into their agricultural policies. For instance, in the Global South, Costa Rica's Pago por Servicios Ambientales programme (Zbinden and Lee, Reference Zbinden and Lee2005), Mexico's Programas de Servicios Ambientales that provide payments to landowners for ecosystem services (Alix-Garcia et al., Reference Alix-Garcia, Shapiro and Sims2012), South Africa's Working for Water Programme that controls invasive alien plants (Van Wilgen et al., Reference Van Wilgen, Forsyth, Le Maitre, Wannenburgh, Kotzé, van den Berg and Henderson2012) and China's PESs for watershed conservation (Pan et al., Reference Pan, Xu, Yang and Yu2017) are notable examples. These incentive programmes are often implemented through agri-environmental schemes (AESs)Footnote 4 and payment for ecosystem services (PESs),Footnote 5 where farmers are incentivized or paid to reduce agriculture's impact on water quality and to promote positive environmental effects (McGurk et al., Reference McGurk, Hynes and Thorne2020). However, the voluntary nature of these programmes often leads to low participation rates and limited effectiveness in achieving desired environmental outcomes (Matzdorf and Lorenz, Reference Matzdorf and Lorenz2010; European Court of Auditors [ECA], 2011).
To address these challenges, ECA (2011) recommends providing sufficient financial support and tailoring programmes to local and regional conditions. However, simply tailoring programmes to local conditions is not sufficient. Farmers are diverse, with varying management strategies, cost structures and subcultures (Brodt et al., Reference Brodt, Klonsky and Tourte2006; Howden et al., Reference Howden, Vanclay, Lemerle and Kent1998; Jaeck and Lifran, Reference Jaeck and Lifran2014). As a result, they have different incentives and goals for participating in agNPS pollution control schemes. Despite these differences, current AES and PES contracts fail to account for farmer heterogeneity. In addition, the traditional approach of treating farmers as a homogeneous group and providing flat-rate payments based on average income losses is outdated. This approach overcompensates farmers with lower-than-average compliance costs while deterring those with higher-than-average costs (Latacz-Lohmann and Breustedt, Reference Latacz-Lohmann and Breustedt2019). A more nuanced approach that recognizes and addresses farmer heterogeneity is necessary to improve the effectiveness of AES and PES programmes.
Accordingly, understanding how these differences drive farmers’ compensation requirements is fundamental to designing appropriate and more acceptable water-quality improvement-related AESs and PESs in South Africa and the Global South to successfully regulate agNPS pollution. Hence, the purpose of this paper is to investigate the extent to which the adoption of more sustainable farming practices under financial compensation can contribute to reducing agNPS pollution in the Limpopo River Basin (LRB) of South Africa. We target sustainable practices that range from reduced use of fertilizers and pesticides during production to the construction of ecological ditches and waste recovery initiatives, all of which contribute to reducing releases of harmful pollutants.
To this end, we first determine the appropriate compensation that farmers are willing to accept to adopt the aforementioned farming practices. Second, we determine the socio-economic and attitudinal factors that drive farmers’ willingness to control agNPS pollution. Finally, we explore unobserved heterogeneity among farmers in the sample. With these steps, we hope to provide new insights to enrich the efficient design of tailored water quality improvement-related AESs and PESs that promote full participation, persistent environmental benefits and positive externalities that benefit farmers and society.
This is the first study, to the best of our knowledge, that uses a choice experiment to elicit farmers’ compensation requirements to control agNPS pollution in South Africa. We focused on the LRB for several reasons: (1) despite its importance for agriculture, eco-tourism, mining and industry, which contribute to promoting employment, income and poverty alleviation, the LRB is one of the most polluted river systems in South Africa (Marr et al., Reference Marr, Mohlala and Swemmer2017); (2) remedying this pollution is a key objective of the government to restore and protect the basin under the National Water Resource Strategy 2 (NWRS 2) (2013); (3) there is a lack of empirical research in South Africa on the use of monetary incentives to motivate farmers to control agNPS pollution. This hinders the development of cost-effective and proactive policies to regulate agNPS pollution in the agricultural sector; and (4) studying the LRB offers a unique perspective on policy regulations for agNPS pollution in the context of extreme water scarcity and an emerging economy context.
Our study makes two key contributions. First, it highlights the importance of incorporating farmers’ behavioural traits and diverse preferences into the design and implementation of AES and PES contracts aimed at improving water quality. Secondly, beyond its academic contributions, our study provides valuable insights and policy guidance for developing cost-effective and proactive strategies to address agNPS pollution, ultimately informing policies in the agricultural sector.
The rest of the paper is structured as follows. Section 2 provides a review of the related literature, and section 3 provides a brief description of the study area. Section 4 comprises the methodology. The results and discussion appear in sections 5 and 6 concludes, provides policy implications and discusses the limitations of the study and directions for future research.
2. Related literature
The literature shows that farmers’ participation in these water quality improvement-related AESs and PESs programmes is highly heterogeneous. While some farmers are willing to accept monetary compensation to participate in these programmes (Li et al., Reference Li, Liu, Yan, Fan and Zhao2019; McGurk et al., Reference McGurk, Hynes and Thorne2020), other farmers are hesitant, indicating strong preferences for the status quo (Beharry-Borg et al., Reference Beharry-Borg, Smart, Termansen and Hubacek2013; Christensen et al., Reference Christensen, Pedersen, Nielsen, Mørkbak, Hasler and Denver2011). Some farmers put emphasis on the regulatory agency that would implement and monitor the programme and the amount of land that would be committed to it (Kreye et al., Reference Kreye, Pienaar, Soto and Adams2017). Moreover, some farmers required greater financial incentives to participate in schemes with longer contract durations, less flexibility and/or higher levels of paperwork (Ruto and Garrod, Reference Ruto and Garrod2009). Finally, fixed transaction costs are perceived as a significant barrier to farmers’ interest in scheme contracts, particularly for small farmers. Therefore, lump-sum payments that are yearly are perceived to increase participation rates (Ducos et al., Reference Ducos, Dupraz and Bonnieux2009).
We present an overview of key-related empirical studies and their findings from around the world, especially those from the Global South, in table A1 of the online appendix. The reviewed studies have contributed significantly to this literature. However, there are notable gaps that this study aims to address. First, none of the reviewed studies are from South Africa, which provides a unique context for investigating water quality regulation in an emerging economy context. South Africa's growing food production and water use demands necessitate efficient agricultural systems, making local research essential. Secondly, research is needed on tailored incentives that (1) account for farmers’ behavioural traits and preference differences, (2) recognize variations in farm management practices and cost structures and (3) encompass diverse farming practices beyond single-crop systems. Our study addresses this by focusing on farmers growing multiple crops, including vegetables, spices, sweet potatoes and maize. This diversity is crucial, as these crops are more susceptible to fertilizer and pesticide leaching due to frequent cultivation and shorter growing cycles (Di and Cameron, Reference Di and Cameron2002). By considering diverse crops, our study provides a more nuanced understanding of farmers’ decision-making and introduces flexibility to the options available to individual farmers.
Our study fills these critical gaps and lays the groundwork for incorporating willingness to accept (WTA) compensation into agricultural decision-making. By providing a more nuanced understanding of the strategies to control agNPS pollution and a wider range of farm-level mitigation practices, this research aims to inform evidence-based solutions for sustainable agriculture and water management. To achieve this, we designed a hypothetical market scenario where farmers agreed to adopt sustainable farming practices, such as reducing fertilizer and pesticide use and recovering agricultural waste, among other initiatives, in exchange for compensation.
3. An overview of the study area
The LRB, as shown in figure A1 of the online appendix, covers an area of 416,296 square km. It is the second longest (approximately 1,750 km or 1,087 miles) and the most significant river in southern Africa (Nakayama, Reference Nakayama2003). The basin is shared by four countries – Botswana (19 per cent), Zimbabwe (15 per cent), Mozambique (21 per cent) and South Africa (45 per cent). South Africa’s largest share, which predominantly lies in the Limpopo Province, spans approximately 184,150 square km (Limpopo Basin Permanent Technical Committee, 2010). The river is important because of several economic and ecological factors. First, the river plays a crucial role in supporting agriculture and livestock farming in the region. Many communities rely on the river for irrigation, and the fertile soils around it support the cultivation of many crops, including cotton, maize and citrus fruits. Secondly, the river is significant for wildlife conservation. Kruger National Park, one of the largest game reserves in Africa and home to the iconic African Big Five (lion, leopard, elephant, rhinoceros and Cape buffalo), is found in this area. Even though the river is crucial to the province's people, wildlife and economy, it faces significant threats from climate change and water pollution. In this regard, two farming communities in the basin, Folovhodwe and Tshiombo in the Vhembe District Municipality of the province, were chosen to understand how to address the issues of water scarcity and quality and to ensure the river's long-term sustainability. Folovhodwe and Tshiombo are located on important tributaries of the Limpopo River. The Nwanedi River, which houses the Nwanedi Irrigation Scheme, passes through Folovhodwe, while the Tshiombo Irrigation Scheme, one of the largest in the province, is at the western end of the Tshiombo Valley on the south bank of the Mutale River (Lahiff, Reference Lahiff1997). Agricultural activities predominantly include growing vegetables, bananas, citrus, maize, melons and peanuts, as well as poultry and livestock production.
4. Methodology
4.1. Model specification
The discrete choice experiment (DCE)Footnote 6 methodology has its theoretical foundation in Lancaster's model of consumer choice (Lancaster, Reference Lancaster1966) and its econometric basis in random utility theory (Luce, Reference Luce1959; McFadden, Reference McFadden and Zarembka1974). Based on random utility theory, farmers choose the alternative that provides them with the highest expected utility associated with the choice attributes. The 
$i$th farmer's utility of the 
$j$th alternative associated with choosing the scheme's intervention is given as follows:
\begin{equation}{U_{ij}} = {\beta _i}{x_{ij}} + {\varepsilon _{ij}} = {\text{ }}{V_{ij{\text{ }}}} + {\varepsilon _{ij}},{\text{ }}\end{equation} where 
${\beta _i}$ is a vector of individual-specific coefficients and 
${x_{ij}}$ is a vector of observed agNPS pollution control attributes related to the 
$i$th farmer and the 
$j$th alternative. The utility (
${U_{ij}}$) of a choice set, thus, comprises a deterministic part (
${V_{ij}}$) and a random component (
${\varepsilon _{ij}}$).
Assuming the random components in equation (1) are independent and identically distributed (IID), with a Gumbel (0,1) distribution (Greene, Reference Greene1997), then the conditional logit model (CLM) of the probability of the 
$i$th farmer choosing the 
$j$th alternative is given as follows:
\begin{equation*}{P_{ij}} = \frac{{{\text{exp}}\left( {\lambda {V_{ij}}} \right)}}{{\mathop \sum \nolimits_{l = 1}^J {\text{exp}}(\lambda {V_{il}})}},\end{equation*} where 
$J$ is the set of available agNPS pollution control alternatives and 
$\lambda = 1$.
Following Hole (Reference Hole2007), the log-likelihood function of the CLM is given as follows:
\begin{equation*}{{log}}\mathcal{L} = \mathop{\sum}\limits_{i = 1}^N \mathop{\sum}\limits_{j = 1}^M {y_{ij}}ln\left[ {\frac{{{\text{exp}}\left( {{V_{ij}}} \right)}}{{{\text{exp}}\left( {{V_{il}}} \right)}}} \right]\!\!,\end{equation*} where 
${y_{ij}}$ is an indicator variable that is equal to 1 if the 
$j$th alternative is chosen by farmer 
$i$ and zero otherwise, 
$N$ is the total number of farmers (
$i = 1,{\text{ }}2, \ldots ,{\text{ }}N)$ and 
$M$ indicates the 
$j$th alternative in a choice set of the farmer (
$j = 1,{\text{ }}2, \ldots ,{\text{ }}M$).
However, the CLM's strict assumption of the independence of irrelevant alternativesFootnote 7 makes other advanced models like the random parameter logit model (RPLM) and the latent class logit model (LCM) that relax this assumption more desirable. Furthermore, the CLM assumes that preferences are homogenous across respondents (Mariel et al., Reference Mariel, Hoyos, Meyerhoff, Czajkowski, Dekker, Glenk, Jacobsen, Liebe, Olsen, Sagebiel and Thiene2021). However, farmers are heterogeneous due to various factors, including socio-economic characteristics and cost structures, among others. Ignoring this heterogeneity can lead to biased estimates (Greene, Reference Greene1997). Accordingly, we employed the RPLM and the LCM to address these limitations. By using both approaches, we gain a more complete understanding of the underlying preferences and stated behaviour, capturing both continuous and discrete heterogeneity. We note, however, that our use of these models is exploratory, aiming to provide a richer understanding of preference heterogeneity, rather than validating or comparing these models.
The RPLM accounts for preference heterogeneity by allowing utility parameters to vary randomly and continuously over individuals (McFadden and Train, Reference McFadden and Train2000), while the LCM postulates a discrete distribution of taste in which individuals are intrinsically put into a number of segments (or classes). The RPLM relaxes the assumptions and limitations of the CLM by allowing random taste variation within a sample according to a specified distribution (McFadden and Train, Reference McFadden and Train2000). This approach recognizes that individuals may have different sensitivities to our attributes. By estimating the distribution of these random parameters, we can capture the variation in preferences across the population. Thus, following Hole (Reference Hole2007), the unconditional probability that the 
$i$th farmer chooses the 
$j{\text{th}}$ alternative at choice occasion 
$t$ is given as follows:
\begin{equation*}{P_{ijt}}\left( \phi \right) = \smallint\frac{{{\text{exp}}\left( {\beta _i^{'}{x_{ijt}}} \right)}}{{\mathop {\sum} \limits _{l = 1}^L{\text{exp}}\left( {\beta _i^{'}{x_{ilt}}} \right)}}f\left( {{\beta _i}|\phi } \right)d{\beta _i} ,\end{equation*} where 
$f\left( {\beta |\phi } \right)$ is the probability density function for 
$\beta $, which is a vector of parameters specific to the 
$i$th farmer. It represents the farmer's tastes and varies across farmers. 
$\phi $ is a vector of parameters that describe the density of the distribution of the individual-specific parameters 
$\beta $ with mean 
$b$ and covariance 
$W$. The parameters of the RPLM are estimated by maximizing the simulated log-likelihood function that is given as follows:
\begin{equation*}{\text{Sim}}LL = \mathop{\sum}\limits_{i = 1}^I {ln}\left\{ {\frac{1}{R}\mathop{\sum}\limits_{r = 1}^R \left( {\mathop{\prod}\limits_{t = 1}^T \mathop{\prod}\limits_{j = 1}^J {{\left[ {\frac{{{\text{exp}}\left( {\beta _i^{r'}{x_{ijt}}} \right)}}{{\mathop{\sum}\nolimits_{l = 1}^L{\text{exp}}\left( {\beta _i^{r'}{x_{ilt}}} \right)}}} \right]}^{{f_{ijt}}}}} \right)} \right\},\end{equation*} where 
$R$ is the number of replications, 
$\beta _i^{r'}$is the 
$r{\text{th}}$ draw for the 
$i$th farmer from distribution of 
$\beta $ and 
${f_{ijt}}$ is a dummy equal to 1 if farmer 
$i{\text{ }}$ chooses alternative 
$j$ at choice occasion 
$t$ and zero otherwise.
However, given that the RPLM is unable to identify segmentations (classes), the LCM is further estimated in this regard. The LCM can allow decision makers to explicitly account for unobserved heterogeneity. It does this by implicitly grouping individuals that exhibit similar (unobservable to the analyst) preferences into specific classes or sub-groups based on some probabilities (Boxall and Adamowicz, Reference Boxall and Adamowicz2002). These sub-groupings then make it possible for the different concerns of the different groups of farmers to be addressed with policies that are tailored to each group. In the context of policy, accounting for heterogeneity offers several benefits. It provides a more comprehensive understanding of how policies affect different groups, enabling policymakers to (1) assess distributional consequences and other policy impacts, (2) design policies that address equity concerns and diverse behaviours and (3) provide better insights into policy outcomes (Greene, Reference Greene2011; Garrod et al., Reference Garrod, Ruto, Willis and Powe2012).
Following Swait (Reference Swait1994), the utility function for the LCM of the 
$i$th farmer's choice among 
$J$ alternatives, given that the farmer belongs to class 
$c = 1, \ldots ,{\text{ }}C$, is expressed as follows:
\begin{equation}{U_{ij/c}} = {\beta _c}{x_{ij}} + {\varepsilon _{ij/c}},{\text{ }}\end{equation} where 
${x_{ij}}$ is a vector of attributes associated with alternative 
$j$, 
${\beta _c}$ is a class-specific parameter vector associated with the vector 
${x_{ij}}$ and 
${\varepsilon _{ij/c}}$ represents the random variations for the 
$i$th farmer. If the error terms are IID and follow a Type 1 extreme value distribution across classes and farmers, the probability that the 
$i$th farmer belongs to class 
$c$ and selects alternative 
$j$ is given by the following equation:
\begin{equation*}{P_{ij/c}} = \frac{{{\text{exp}}\left( {{\beta _c}{x_{ij}}} \right)}}{{\mathop{\sum}\nolimits_J{\text{exp}}\left( {{\beta _c}{x_{ij}}} \right)}}.\end{equation*} The joint probability of farmer 
$i$ belonging to class 
$c$ and selecting alternative 
$j$ is given as follows:
\begin{equation*}{P_{iJC}} = {P_{ij/c}}*{P_{ic}},\text{ where } {P_{ic}} = \frac{{{\text{exp}}\left( {\delta {Z_i}} \right)}}{{\mathop{\sum}\nolimits_C{\text{exp}}\left( {\delta {Z_i}} \right)}},\end{equation*} with 
${Z_i}$ being a vector of the class-specific parameters and 
$\delta $ being a scale factor 
$ = 1$. Accordingly, the marginal probability of observing the 
$i$th farmer in class 
$c$ choosing alternative 
$j$ is expressed as follows:
\begin{equation*}{P_{ij}} = \mathop{\sum}\limits_{c = 1}^C \left[ {\frac{{{\text{exp}}\left( {{\beta _c}{x_{ij}}} \right)}}{{\mathop{\sum}\nolimits_J{\text{exp}}\left( {{\beta _c}{x_{ij}}} \right)}}} \right]\left[ {\frac{{{\text{exp}}\left( {\delta {Z_i}} \right)}}{{\mathop{\sum}\nolimits_C{\text{exp}}\left( {\delta {Z_i}} \right)}}} \right]\!\!.\end{equation*} Equation (2) implies that the probability of selecting the 
$j$th alternative is equal to the sum over all latent classes 
$c$ of the class-specific membership model, conditional on (
${P_{ij/c}}$) multiplied by the probability of belonging to that class (
${P_{ic}}$). The log-likelihood function to obtain the parameters 
$\delta $ and 
${\beta _c}$ is given as follows:
\begin{equation*}log\mathcal{L} = \mathop \sum \limits_{n = 1}^N \mathop \sum \limits_{j = 1}^J {y_{ij}}ln{\text{ }}\left( {\mathop \sum \limits_{c = 1}^C {P_{ij/c}}*{P_{ic}}} \right)\!\!,\end{equation*} where 
$J$ is the total number of alternatives and 
${y_{ij}}$ is the observed frequency of choice of alternative 
$j$ by the 
$i$th farmer. All other indicators have their usual meaning.
To estimate our LCM, we need to determine the suitable number of latent classes that balance fit, interpretability, simplicity and theoretical consistency (Boxall and Adamowicz, Reference Boxall and Adamowicz2002). Additionally, to account for inattention bias, we allow for a random response share, assuming some choices are made randomly (Malone and Lusk, Reference Malone and Lusk2018; Lagerkvist et al., Reference Lagerkvist, Edenbrandt, Tibbelin and Wahlstedt2020). Here, the probability of class membership for one class is determined completely by the random utility term (Malone and Lusk, Reference Malone and Lusk2018), with systematic utility restricted to zero for all attribute parameters (Lagerkvist et al., Reference Lagerkvist, Edenbrandt, Tibbelin and Wahlstedt2020). We then repeatedly estimate the same LCM with different numbers of classes, including one class of random choice, and select the best model based on the minimum values of the Akaike information criterion (AIC), Bayesian information criterion (BIC) and consistent AIC (CAIC) (Killian et al., Reference Killian, Cimino, Weller and Seo2019).
Once the parameter estimates of the LCM are obtained, the marginal WTA (mWTA) to gauge the minimum WTA requirement for farmers to forgo some undesirable farm management practices to deliver some improvements in water quality is derived as follows:
\begin{equation*}WT{A_{i,k,l}} = \mathop \sum \limits_{c = 1}^C {P_{ic}}\left( {\frac{{ - {\beta _{c,{\text{ }}attributes,{\text{ }}asc}}}}{{{\beta _{c,{\text{ }}compay}}}}} \right)\!\!,\end{equation*} where 
${P_{ic}}$ is the estimated matrix of individual 
$n$-specific probabilities of segment membership, and 
$\left( {\frac{{ - {\beta _{c,{\text{ }}attributes,asc}}}}{{{\beta _{c,{\text{ }}compay}}}}} \right)$ is the ratio of implicit price for the attribute being valued, relative to the probability of the status quo option. In addition, following Hanemann (Reference Hanemann1984), we estimate the compensating surplus (CS; 
$CS$) for different water quality improvement scenarios based on a combination of some attributes,
\begin{equation*}CS = - \frac{1}{{{\beta _{compay}}}}{\text{ }}\left[ {{\text{ln}}\left( {\mathop{\sum}\limits_i {\text{exp}}\left( {V_i^1} \right)} \right) - {\text{ln}}\left( {\mathop{\sum}\limits_i {\text{exp}}\left( {V_i^0} \right)} \right)} \right]\!\!,\end{equation*} where 
${\beta _{compay}}$ is the coefficient of the compensation attribute. It captures the marginal utility of income. 
$V_i^0$ and 
$V_i^1$ represent the 
$i$th farmer's indirect utility functions before and after the change under consideration.
Finally, Johnston et al. (Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa and Tourangeau2017) explicitly recommend accounting for attribute non-attendance (ANA) as a potential behavioural anomaly in DCE studies, where respondents ignore or do not consider one or more attributes when making their choices (Hensher et al., Reference Hensher, Rose and Greene2005). Accordingly, we estimate the endogenous attribute attendance (EAA) model of Hole (Reference Hole2011) and Hole et al. (Reference Hole, Kolstad and Gyrd-Hansen2013). In this model, respondents are assumed to choose a subset 
${C_q}$ from a total of 
$K$ attributes to consider when choosing an alternative. The total number of attribute subsets is given by 
$Q = {2^K}$, which includes the set in which all attributes are included (
${C_Q}$) and the empty set in which the respondents discard all the information about the alternatives (
${C_1}$). The former corresponds to the conventional assumption that the decision makers make use of all the available information on the alternatives when making a choice, while the latter implies that the choice process is random. Conditional on the choice of attribute subset 
${C_q}$, the utility that individual 
$i$ derives from choosing alternative 
$j$ on choice occasion 
$t$ is given by 
${U_{ijt}} = \mathop \sum \limits_{k \in {C_q}} X_{ijt}^k{\beta ^k} + {\varepsilon _{ijt}}$, where 
$X_{ijt}^k$ represents the value of attribute 
$k$ relating to alternative 
$j{\text{ }}$on choice occasion 
$t$, 
${\beta ^k}$ is the preference weight given to that attribute and 
${\varepsilon _{ijt}}$ is a random term which is assumed to be IID extreme value. Accordingly, the probability that farmer 
$i$ chooses alternative 
$j$ on choice occasion 
$t$ conditional on the choice of attribute subset 
${C_q}$ is given by the following equation:
\begin{equation}{\text{Pr}}\left( {choic{e_{it}} = j|{C_q}} \right) \frac{{{\text{exp}}\left( {X_{ijt}^k{\beta ^k}} \right)}}{{\mathop {\sum} \nolimits_{j = 1}^J \exp\left( {X_{ijt}^k{\beta ^k}} \right)}}.\end{equation} The probability that farmer 
$i$ takes attribute 
$k$ into account is specified as 
${\text{exp}}\left( {\delta _k^{'}{Z_{ik}}} \right)/\left[ {1 + {\text{exp}}\left( {\delta _k^{'}{Z_{ik}}} \right)} \right]$, where 
${Z_{ik}}$ is of individual-level observed characteristics and 
${\delta _k}$ is a vector of parameters to be estimated (Ding and Abdulai, Reference Ding and Abdulai2018). Assuming that these probabilities are independent over attributes, the probability of choosing attribute subset 
${C_q}$ is given by the following equation:
\begin{equation}{P_{i{C_q}}} = \mathop{\prod}\limits_{k \in {C_q}} \frac{{\exp\left( {\delta _k^{'}{Z_{ik}}} \right)}}{{1 + \exp\left( {\delta _k^{'}{Z_{ik}}} \right)}}\mathop{\prod}\limits_{k \notin {C_q}} \frac{1}{{1 + \exp\left( {\delta _k^{'}{Z_{ik}}} \right)}}.{\text{ }}\end{equation}Combining equations (3) and (4), the unconditional probability of the observed sequence of choices is
\begin{equation*}P_i^{EAA} = \mathop \sum \limits_{q = 1}^Q {P_{i{C_q}}} \times \mathop \prod \limits_{t = 1}^T \mathop \prod \limits_{j = 1}^J {\text{Pr}}{(choic{e_{it}} = j|{C_q})^{{y_{ijt}}}},\end{equation*} where 
${y_{ijt}}{\text{ }}$is equal to 1 if farmer 
$i$ chooses alternative 
$j$ on choice occasion 
$t$ and zero otherwise. The 
$\beta $ and 
$\delta $ parameters are estimated jointly by maximizing the log-likelihood function:
\begin{equation*}\mathcal{L}{\mathcal{L}^{EAA}} = \mathop \sum \limits_{i = 1}^I {{ln}}P_i^{EAA}.\end{equation*}4.2. Design of the choice experiment
4.2.1. Attributes, levels and validation
Based on the literature, interviews with experts in the region and focus group discussions, we identified and used seven attributes: (1) fertilizer reduction, (2) pesticide reduction, (3) agricultural waste recovery initiatives, (4) construction of ecological ditches, specifically vegetative strips, (5) duration of the programme in years, (6) monitoring for compliance and (7) the compensation payment in United States dollars per hectare per year – and their respective levels. In addition to the DCE variables, socio-demographic and attitudinal characteristics of representative farmers, like age, education, farm and off-farm income, and farming experience, among others, were collected. In table 1, we present our attributes, their descriptions and levels.
Attributes, their descriptions and levels
Notes: ZAR is the South African Rand, and US$ is the United States dollar. The exchange rate during the survey was 1 US{{footpara}}amp;#x00A0;= 15 ZAR (in April 2021).
To ascertain the validity of our questionnaire and choice sets, two focus group discussions were conducted – one with farmers and extension service officers and the second with stakeholders and industry experts from the Department of Agriculture, Forestry, and Fisheries and the De Beers Group.Footnote 8 The discussions centred on the purpose of the research, input prices and farmers’ revenues, and how to make our questionnaire understandable to farmers. The discussions resulted in an upward revision of all the levels of compensation, from their initial levels of US$200, US$600 and US$1,400 per hectare per year to their current levels. Overall, the participants validated our choice attributes and their levels.
4.3. Experiment and survey design
A fractional factorial design was generated using NGENE 1.2.1 software, resulting in 72 choice sets that were ‘blocked’ into eight blocks of nine choice sets. Each farmer was then randomly assigned one of the eight versions of our questionnaire and asked to make nine choices during the face-to-face interview. Our questionnaire was structured to include cheap talk and consequentiality checks (see online appendix A1) to minimize hypothetical bias and improve the validity and reliability of our results (Johnston et al., Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa and Tourangeau2017). Recent advances in choice literature have identified conditions for stated preference surveys to be incentive-compatible, meaning respondents are incentivized to truthfully reveal their preferences rather than misrepresent them (Carson et al., Reference Carson, Groves and List2014). Vossler et al. (Reference Vossler, Doyon and Rondeau2012) identify conditions for the incentive compatibility of the elicitation strategy, including when a single binary choice question is used in the increasingly popular repeated binary DCE – one control alternative and the status quo (Vossler and Evans, Reference Vossler and Evans2009; Carson et al., Reference Carson, Groves and List2014). However, each of these formats have their advantages and disadvantages, including differences in incentive properties (Carson et al., Reference Carson, Groves and List2014; Collins and Vossler, Reference Collins and Vossler2009).
Given the objective of estimating farmers’ WTA compensation to control agNPS, which involves various sustainable farming practices, we adopted the traditional two-alternatives and one status quo option approach. This approach, when compared to the binary approach, has several advantages. First, with two alternatives and a status quo option, each choice set provides more information than a binary choice question, allowing for more efficient data collection, as fewer choice sets are needed to achieve the same level of precision, reducing respondent burden and data collection costs (Rose and Bliemer, Reference Rose and Bliemer2013). Secondly, our approach provides a better representation of real-world choices, where individuals often face multiple alternatives rather than binary choices (Louviere et al., Reference Louviere, Hensher DA and Swait JD2000). This complexity is better captured in our study, resulting in more realistic choice scenarios. Accordingly, each of our choice sets consisted of three alternatives – two agNPS pollution control alternatives and the status quo alternative. Farmers who chose the control alternatives desired improvements in water quality, whereas those who chose the status quo were unwilling to undertake actions to improve water quality.
An example of one of our choice cards is presented in online appendix A2. We chose face-to-face interviews because this is the gold standard and most effective survey method for stated preference studies in developing countries (Bennett and Birol, Reference Bennett, Birol, Bennett and Birol2010). This way, our enumerators were able to explain the purpose of the study and choice tasks objectively to farmers in their own language. Before commencing the interviews, our enumerators were taken through a thorough two-day training exercise to master the questionnaire and interview techniques. To ensure that our farmers clearly understood the attributes and their levels, enumerators used a detailed album with coloured pictures of attributes. This visual aid helped to explain attributes in a clear and concise manner, reducing interviewer bias and overcoming language barriers (Bennett and Birol, Reference Bennett, Birol, Bennett and Birol2010).
4.4. Sampling and data
A two-stage sampling approach was used in selecting the study area and the farmers. In the first stage, the Folovhodwe and Tshiombo farming communities were purposively selected not only because these are agrarian hubs in the Vhembe District of the Limpopo Province, but also because both communities are located near important tributaries of the Limpopo River, as mentioned earlier. In addition, the frequent droughts in the area, coupled with the anthropogenic activities, as well as the area having a large number of documented and undocumented farmers with well-organized farmer associations that are readily accessible and available for research, informed our choice of these two farming communities. In the second stage, a simple random sampling technique was applied to select farmers from a population of both documented and undocumented farmers on the farming blocks in the study area. There are about 74,073 agricultural households in the Vhembe District (Statistics South Africa, 2022). Each farmer on one of the eight blocks, whether documented or undocumented, had an equal chance of being selected. In coordination with farmers’ leaders and extension services officers, farmers were informed of the impending interviews. Enumerators then went to farmers on their farms and conducted one-on-one interviews using our questionnaires. Data collection lasted one month, between March and April 2021.
A total of 559 interviews were conducted. However, at the data entry phase, seven questionnaires were rejected because of incomplete or inconsistent information. This high success rate of questionnaires completed was a result of the training the enumerators received, and the thorough checks put in place to scrutinize each questionnaire before accepting it. In these checks, if any errors or incomplete fields were detected, the enumerators were requested to go back to the respective farmer(s) to complete the questionnaire before we accepted it. Our response rate was then compared to the minimum sample size required for adequate power, following Cochran (Reference Cochran1963) (see online appendix A3). We then concluded that our sample of 552 farmers is adequate for power analysis.
5. Results and discussions
5.1. Descriptive statistics
The descriptive statistics of our socio-demographic and attitudinal characteristics are presented in table A2 of the online appendix. These statistics show that females constitute 55 per cent of our sample. The 2011 Census also shows that females constitute 54 per cent of the population of Folovhodwe and Tshiombo (Census, 2011), confirming this statistic. The average age of our sample is 51. With regard to education, 94 per cent are literate. The average farming experience is 14 years. Farmers with farm income greater than US$2,733Footnote 9 (in April 2021) after harvest constitute 45 per cent of our sample, while 69 per cent have off-farm income activities. Seventy-six per cent have secured tenure rights to their farmlands. The average farm size is 3.28 hectares. Some 55 per cent of farmers have no awareness of agNPS pollution, but 98 per cent believe that they stand to benefit if water quality is improved. Ninety-three per cent own the farms they work on, while the rest rent or work on leased lands. On farm type, 93 per cent have diversified farms (grow multiple crops and rear different animals). Furthermore, 89 per cent use surface water for irrigation, while 62 per cent are members of a cooperative. To ascertain the representativeness of our sample, we compared our statistics to those of the Vhembe District for the 2011 Population Census. In most cases, these statistics matched, demonstrating that our sample is representative of the population.
5.2. CLM and RPLM
The results of our CLM and RPLMs are presented in table A3 of the online appendix. Both models are estimated such that the probability of selecting any alternative is a function of the different levels of the attributes and the alternative-specific constant for the status quo option (ASCsq). ASCsq is a dummy variable. A positive parameter estimate of the ASCsq indicates a preference for the status quo, suggesting that, on average, farmers prefer the status quo option over the scheme's alternatives (Hensher et al., Reference Hensher, Rose and Greene2015). Conversely, a negative parameter estimate implies a lack of preference for the status quo (Hensher et al., Reference Hensher, Rose and Greene2015).
5.2.1. The CLM
Our results show that the coefficient of ASCsq is negative and highly significant. This indicates that our farmers prefer the scheme's alternatives (agNPS pollution control alternatives) and are willing to alter their undesirable farm management practices to improve water quality. In addition, the coefficient of compensation payment is positive and highly significant, suggesting that an increase in the compensation payment is more likely to lead to an increase in the probability of choosing the agNPS pollution control alternatives, all else being constant. Furthermore, our analysis shows that the coefficients of all the other choice attributes have the expected a priori signs (negative) and are highly significant, except that the different levels of agricultural waste recovery have the right signs but are insignificant. This result suggests that, in comparison with their status quo levels, fertilizer and pesticide reductions, ecological ditches, monitoring and the duration of the programme decrease the probability of accepting the agNPS pollution control programme, all else constant. According to Hanemann (Reference Hanemann1991), such results indicate that respondents derive disutility or disprefer the attributes. Therefore, they require higher compensation to accept higher levels of the attributes.
5.2.2 The RPLM
We estimated the mixlogit of Hole (Reference Hole2007) using 500 Halton draws for model convergence because of their higher efficiency compared to random draws (Campbell, Reference Campbell2007; Hole, Reference Hole2011). In addition, based on the literature (McFadden, Reference McFadden and Zarembka1974; Li et al., Reference Li, Liu, Yan, Fan and Zhao2019), compensation payment is fixed, while the choice attributes are set to follow a normal distribution. Two different RPLMs are estimated – RPLM I used only the scheme's attributes, while RPLM II adds some interaction terms. Specifically, we examine interactions between the ASCsq and individual characteristics, such as gender, age and so on. Significant estimates of these interaction terms indicate that gender, age, education and farming experience, among others, reveal significant sources of preference heterogeneity (Pan et al., Reference Pan, Zhou, Zhang and Zhang2016). RPLM II shows identical coefficients to RPLM I, indicating the robustness of our results. The RPLMs and the CLM show negative coefficients of ASCsq, indicating a similar interpretation to those of CLMs. The standard deviation (SD) of the ASCsq, which represents the heterogeneity in farmers’ preferences for the status quo option (Hole, Reference Hole2007), is significant in both RPLMs. This suggests that some farmers have a strong preference for the status quo, while others have a weaker preference or even a dis-preference (McFadden and Train, Reference McFadden and Train2000). Furthermore, the mean coefficient of compensation payment is positive and highly significant in both RPLMs, indicating consistency with the CLM results that compensation payments incentivized farmers and increased their willingness to alter their undesirable farming practices to improve water quality. This finding is consistent with most AES and PES studies (Broch and Vedel, Reference Broch and Vedel2012; Beharry-Borg et al., Reference Beharry-Borg, Smart, Termansen and Hubacek2013; Li et al., Reference Li, Liu, Yan, Fan and Zhao2019).
As with the CLM, all the other choice attributes in both RPLMs are negative and highly significant, except for those of waste recovery. Like the CLM, here too the explanations and implications are similar. In addition, the SD for the different levels of fertilizer and pesticide reductions, ecological ditches, duration of the programme and monitoring are significant, confirming substantial farmer heterogeneity in our data. In both RPLMs, 100 per cent waste recovery is insignificant, but its SD is significant for RPLM II, indicating an insignificant and consistent effect of waste recovery on utility functions.
Finally, in RPLM II, the mean of the interaction of education, farming experience, farm income and security of tenure is negative and significant at the 1 per cent level. This suggests farmers who are literate, have over 14 years of farming experience, earn over US$2,733 and have security of tenure are less likely to choose the status quo option relative to their colleagues who are uneducated, have less farming experience, earn lower incomes and lack secure tenure, all else being constant. These findings are consistent with the existing literature, which highlights the importance of factors such as age, education, farming experience, farm income and security of tenure, among others, as influencing or shaping farmers’ decisions regarding AES and PES adoptions.
First, education increases awareness and understanding of these schemes and enhances farmers’ ability to make informed decisions about participation. Higher education levels are generally associated with increased AES and PES adoptions (Pavlis et al., Reference Pavlis, Terkenli, Kristensen, Busck and Cosor2016). Secondly, experienced farmers may find it easier to transition to AES and PES participation due to accumulated knowledge (Wilson, Reference Wilson1997). Thirdly, farmers with higher incomes may be more likely to participate in these schemes, as they are better able to absorb the costs associated with scheme participation (Pascual et al., Reference Pascual, Muradian, Rodríguez and Duraiappah2010). Finally, farmers with secure tenure rights are more likely to participate in these schemes to promote environmental conservation, as they have a long-term stake in the land (Besley, Reference Besley1995). However, the mean of the interaction of age is positive and statistically significant at the 5 per cent level. This suggests an increase in the average probability of choosing the status quo option if the farmer is more than 51 years old, all else being constant. This is understandable, as older farmers tend to be reluctant to participate in AESs and PESs (Knowler and Bradshaw, Reference Knowler and Bradshaw2007).
5.3. The LCM
The results of the model selection criteria presented in table A4 of the online appendix show that both BIC and CAIC favour five classes (four preference classes and one random choice class), while AIC favours seven classes (six preference classes and one random choice class). Based on the BIC and CAIC statistics, the five-class latent model was initially estimated, but we realized that some of the model's estimates did not make sense theoretically. For instance, some coefficients of compensation payments were zero in some classes. In this regard, Nylund et al. (Reference Nylund, Asparouhov and Muthén2007) and Weller et al. (Reference Weller, Bowen and Faubert2020) advised that the information criteria should always be evaluated in conjunction with theoretical interpretability. Accordingly, the four-class latent model (FCLM) – three preference classes and one random choice class – is rather estimated. This model had no such theoretically interpretable flaws. Lagerkvist et al. (Reference Lagerkvist, Edenbrandt, Tibbelin and Wahlstedt2020) used this approach when their BIC- and CAIC-preferred model did not produce theoretically consistent results. The results of the FCLM presented in table 2 show that the percentage of farmers strongly associated with Classes 1, 2, 3 and 4 is 49, 22, 26 and 3 per cent, respectively.
Results of the four-class latent logit model with covariates
Notes: Compensation payment attribute is scaled down by 1,000 as its values are huge (in thousands). Standard errors are in parentheses.
Our FCLM demonstrates superior performance compared to our CLM and the RPLMs in predicting our sample's behaviour, exhibiting better model fit and enhanced predictive power, as indicated by the FCLM's lower log-likelihood, AIC and BIC statistics. However, the general aversion of farmers to reducing fertilizer and pesticide use, constructing ecological ditches and being monitored for a period of time to control agNPS pollution in the CLM and the RPLMs is also evident in the FCLM. Furthermore, in addition to the random choice class, three preference classes (low-, moderate- and high-resistance) of farmers are identified based on the compensation amount the respective class requires to alter their status quo farm management practices to improve water quality, or more specifically, the amount the class requires for the ASCsq. More specifically, our analysis shows that the coefficients of ASCsq for Classes 1, 2 and 3 are negative and significant, as expected, just as in our CLM and the RPLMs, indicating the consistency of our farmers’ dislike for the status quo. In addition, the coefficients of compensation payment for these classes are positive and highly significant, also confirming the results of our CLM and the RPLMs and the consistency of farmers’ preferences.
In terms of choice attributes, our analysis for Class 1 shows that fertilizer reduction by 25 per cent, the different levels of pesticide reduction, ecological ditches and monitoring are negative and significant at different levels. This suggests that, in comparison with their status quo levels, these attributes reduce the probability of the farmers of Class 1 participating in the scheme, all else being constant. These farmers are, however, indifferent to fertilizer reduction by 50 per cent, the different levels of agricultural waste recovery and the duration of the programme. Our mWTA estimates across the classes (presented in table 3 in section 5.4) show that this class has the highest compensation requirement to change from the status quo. Thus, we labelled it ‘high-resistance farmers’.
Marginal WTA compensation in US$ per hectare per year by class
Notes: Standard errors are in parentheses. Compensation payment attribute is unscaled in these estimations. The exchange rate used was 1 US{{footpara}}amp;#x00A0;= 15 ZAR (in April 2021). There are two farming cycles per year in the study area; therefore, all payments are to be divided by 2 for one cycle's payment.
In Class 2, all the utility coefficients are negative and statistically significant, suggesting that relative to the status quo levels, these attributes decrease the probability of Class 2 farmers joining the scheme, all else being constant. The waste attribute that was insignificant in the CLM and the RPLMs is significant in this class. This result suggests two possible explanations. First, the LCM captures class-specific effects, where the waste attribute is significant only for this particular class of the population (Greene and Hensher, Reference Greene and Hensher2003). Secondly, the CLM and the RPLMs may be masking the effect of the waste variable by averaging across the entire sample, whereas the LCM reveals its significance for Class 2 (Boxall and Adamowicz, Reference Boxall and Adamowicz2002). The compensation requirement for this class to shift from the status quo is modest relative to the other classes. The class is thus labelled as ‘moderate-resistance farmers’.
The utility coefficients in Class 3 show that the different levels of fertilizer reduction, pesticide reduction by 50 per cent, the 50-meter ecological ditch, the different durations of the programme and partial monitoring are negative and highly significant. This suggests that relative to their status quo levels, these attributes reduce the likelihood of the farmers in Class 3 joining the scheme to control agNPS pollution, all else being constant. Conversely, agricultural waste recovery by 100 per cent is positive and significant, indicating that this attribute increases the likelihood of farmers in this class joining the scheme, all else being constant. Their compensation requirement to shift from the status quo is the least relative to the other classes. Thus, Class 3 is classified as ‘low-resistance farmers’.
On the random choice class (Class 4), our results reveal that only 3 per cent of farmers were categorized as making random decisions, while a significant majority (97 per cent) demonstrated attentive and deliberate decision-making, suggesting that most farmers engaged thoughtfully with our survey. This low rate of inattentive choice responses is attributed to the stringent measures put in place to ensure the quality of our research. First, as explained earlier, our enumerators were trained thoroughly to master the questionnaire and interview techniques. Secondly, to facilitate understanding of the choice questions, a detailed album with coloured pictures of all the attributes and their respective levels was made available to enumerators to help explain the choice sets in both words and pictures. Finally, we included three debrief questionsFootnote 10 in our questionnaire. Our analysis shows that for questions (1), (2) and (3), 96.7 per cent of farmers did not find the choice exercise difficult or burdensome, 98.9 per cent said they fully understood the choice task and 92.8 per cent did not dislike the attribute(s) nor did they want any removed if the programme is implemented.
The result of this class can be attributed to a few reasons: first, being the youngest class and most inexperienced relative to the other classes (see descriptive statistics across the classes in table A5 of the online appendix), this behaviour is not surprising, as the young are always known to rush over issues. Secondly, these farmers may be sceptical about the survey's purpose or may not trust the researchers, leading to careless or inattentive responses (Singer, Reference Singer2003). The responses of this class, nonetheless, contribute to the validity of the preferences of farmers in the other classes, for which attribute choices were not random.
For the covariates of the LCM, the membership coefficients for Class 1 are normalized to zero to allow the remaining coefficients of the model to be identified and interpreted relative to the normalized class (Boxall and Adamowicz, Reference Boxall and Adamowicz2002). The direction of significant coefficients (±) is interpreted as a farmer is more (or less) likely to belong to the respective class than to Class 1. Our analysis shows that older farmers who use surface water sources for irrigation, earn over US$2,733, engage in off-farm activities, have secure land tenure rights and are aware of the threats of agNPS pollution are more likely to have preferences aligned with Class 2. In contrast, experienced farmers who are literate, own their farms and perceive to benefit if water quality is improved are less likely to have their preferences aligned with Class 2. Furthermore, female farmers with diversified farms with farm income above US$2,733 are less likely to have their preferences aligned with Class 3. However, those who are members of a cooperative, with farm sizes greater than 3.28 hectares with secure land tenure rights, are more likely to have their preferences aligned with Class 3. Finally, experienced farmers who are members of a cooperative with secure tenure rights to their farms and are aware of the dangers of agNPS pollution are less likely to have their preferences aligned with Class 4, while farmers with farm income above US$2,733 who have farm sizes greater than 3.28 hectares are more likely to have their preferences aligned with this class.
5.4. mWTA compensation
Our mWTA estimates by class and also the weighted average across the classes are presented in table 3. A comparison of these estimates across the classes shows that where the aversions are strongest, farmers require very high compensation to be induced to accept these attributes, as opined by Hanley et al. (Reference Hanley, Mourato and Wright2001). For instance, although the farmers in Class 1 (high-resistance farmers) have relatively low mWTA values for some of the attributes, they have a strong aversion to external monitoring, so they require US$770 to accept it. This class has the highest compensation requirement for the status quo relative to the other classes. Class 2 farmers (moderate-resistance farmers) are averse to fertilizer and pesticide reductions relative to the other classes. To compensate these farmers adequately to reduce fertilizer, they require US$1,501 and US$2,071 per hectare per year for the 25 and 50 per cent reductions, respectively.
For pesticide reductions, they require US$1,159 and US$1,303 per hectare per year for the 25 and 50 per cent reductions, respectively. Furthermore, relative to the other classes, Class 3 farmers (low-resistance farmers) show a strong aversion towards the duration of the programme. The aversion is, however, strongest with the 10-year programme. To adequately induce these farmers to accept the programme, they require US$1,148 and US$2,238 per hectare per year for the 5- and 10-year programmes, respectively. This finding is consistent with Ruto and Garrod (Reference Ruto and Garrod2009), who found that farmers generally exhibit a preference for shorter-term AES contracts over longer-term ones.
5.5. Welfare measures of different policy scenarios
The CS measures the change in required compensation between the initial situation of agNPS pollution (lower water quality) and subsequent situations (improved water quality) needed to render the farmer indifferent to a change. CS allows policymakers to choose not only the alternatives that provide farmers with the highest utility but also those that provide the needed high water-quality improvement. In table A6 of the online appendix, we evaluated and compared welfare in the base category and the changed situations using all attributes to simulate four water-quality improvement scenarios: high-impact water-quality improvements I (long term – 10 years) and II (short term – 5 years); medium- and low-impact water-quality improvements. Farmers will accept US$6,391 and US$5,093 per hectare per year to implement scenarios I and II, respectively, rather than accept the status quo option in both cases, all else being constant. Furthermore, they will accept US$3,797 and US$3,392 per hectare per year to implement the medium- and low-impact water-quality improvement alternatives, respectively, rather than accept the status quo in both cases, all else being constant. However, the choice of any of these water-quality improvement scenarios is subject to the budget available to the authorities and the kind of water-quality improvement desired.
Overall, our mWTA and CS estimates reflect the behavioural preferences of farmers in AESs and PESs and are in accordance with farmers’ income figures for the Limpopo Province. For instance, Statistics South Africa (2022) puts a smallholder farmer's income per month in the province at US$2,560. In comparison, Hosu and Mushunje (Reference Hosu and Mushunje2013) revealed that the net farm income for the whole farm for any smallholder farmer who keeps 61 livestock units and cultivates an average land size of 2.76 hectares was 104,214 ZAR (approximately US$9,925) in 2013, while the net farm income based on a per hectare basis was 67,757 ZAR (US$6,453), thus confirming our compensation figures.
However, at first glance it may seem our figures are higher compared to global standards. This is because of, first, and of greatest significance, the sensitivity of land issues in South Africa, largely due to the country's complex history of forced land seizures.Footnote 11 This and the debates that were ongoing on land expropriation at the time of our data collection exercise created some misinformation and pushed some farmers to even shy away from our survey, saying the survey was a ploy to source information to take over their lands. These historical antecedents and the land expropriation debates could have influenced farmers’ perceptions and their WTA figures. Secondly, the demand for higher compensation may be driven by other factors, including economic constraints like poverty, inequality and unemployment, as well as their heavy reliance on agriculture for income. Finally, farmers’ distrust of government institutions, limited access to key information and need to manage uncertainty may drive their demand for higher compensation. They likely view this compensation as essential for protecting their well-being, preventing financial losses and ensuring the long-term sustainability of their agricultural operations.
5.6. The endogenous attribute attendance model
The results of our EAA model with the ANA probabilities are presented in table A7 of the online appendix. The analysis in model 1 shows that the probability of waste recovery by 50 per cent is not significant, indicating that it does not explain the probability of ANA. However, all the other attributes are significant at various levels, with lower ANA probabilities for fertilizer reduction by 25 per cent (16 per cent), fertilizer reduction by 50 per cent (17 per cent), pesticide reduction by 25 per cent (11 per cent), pesticide reduction by 50 per cent (16 per cent), agricultural waste recovery by 100 per cent (19 per cent), the 25-meter ecological ditch (15 per cent), 5-year programme duration (27 per cent) and partial monitoring (19 per cent). This suggests that the probability that fertilizer reduction by 25 per cent is ignored in one choice situation is 16 per cent. That of fertilizer reduction by 50 per cent is 17 per cent, and so on. However, the attributes mostly ignored are the 50-meter ecological ditch (68 per cent), followed by the 10-year duration programme (60 per cent) and then external monitoring (55 per cent). These attributes were then excluded from model 2, which presents EAA estimates with jointly estimated ANA probabilities.
Model 2 involves only the attributes with the lowest probability of rejection in model 1. That is, in model 2, the attributes mostly ignored (construction of the 50-meter ecological ditch, the 10-year duration programme and external monitoring) were excluded. These attributes were jointly ignored at 65 per cent when in contest with the other choice attributes. The results of model 2 show that these latter attributes tend to have a low ANA probability of rejection (<18 per cent) when put together in a choice situation. This suggests that these attributes are likely to play a critical role in the agNPS pollution control decision-making process. Furthermore, the probability of rejecting the status quo in both models was high, suggesting that our farmers paid more attention to our choice situations during the survey and that our farmers are interested in shifting away from the status quo situation in the LRB.
6. Conclusion and policy implications
This study investigates farmers’ WTA compensation to control agNPS pollution in the LRB in South Africa using a choice experiment. Conditional logit, random parameter logit and LCMs are used to analyse our data from 552 farmers in the basin. We identified one random choice class (farmers making random responses) and three preference classes of farmers (low-, moderate- and high-resistance) – with dissimilar compensation requirements to alter their farming practices to improve water quality. An endogenous ANA model is also estimated to gauge the extent to which farmers attended to our attributes. The result of this shows the attributes mostly ignored are the 50-meter ecological ditch (68 per cent), followed by the 10-year duration programme (60 per cent), and then external monitoring (55 per cent). These attributes were then excluded from our final model, which included only attributes with low ANA probabilities. These latter attributes tend to have a low ANA probability of rejection (<18 per cent) when put together in a choice situation, suggesting that these attributes are likely to play a critical role in the agNPS pollution control decision-making process.
The study's findings are relevant for policymakers and AES and PES implementers in the Global South and beyond. First, our study underscores the importance of farmers being different with different farm management practices, cost structures and, therefore, different preferences and different compensation requirements to participate in these programmes. Secondly, our inclusion of the ecological ditches provides AES and PES implementers with a cheaper, easier-to-construct and easier-to-maintain technology that is effective at controlling agNPS pollution at the farm level. Finally, our findings provide a cost-effective framework for the formulation of water-quality improvement policies in the agricultural sector to regulate/control agNPS pollution. Since public expenditure decisions may be required if financial compensation is incorporated into agricultural policy for controlling agNPS pollution, our results are of immense benefit to policymakers. Within our framework, they may consider several water-quality improvement options and estimate the costs of each to identify their preferred optimal option for managing water quality while enhancing social welfare in general.
The study has several important implications for policy: (1) our results suggest that farmers are willing to accept monetary compensation to control agNPS pollution. Therefore, we recommend the promotion and use of monetary incentives in the agricultural sector to induce farmers to lessen agriculture's impact on water quality; (2) the divergent behaviour noted in the study is driven by differences in age, education, farming experience, farm income and secured land rights, among other factors. Tailoring AES and PES contracts specifically to farmers’ needs is important for maximum participation and successful environmental outcomes. Therefore, we recommend that any effective strategy or strategies aimed at increasing farmers’ participation in future AESs and PESs must, of necessity, factor in these differences. Finally, AES and PES contracts with extensive/restrictive requirements such as external monitoring and a long contract duration may be effective in ensuring high-quality outcomes but are likely to deter farmers from maximum acceptance of the programme initially. Therefore, we recommend that low-to-moderate requirements be established initially and then farmers be offered the ability to scale up for additional compensation payments.
Despite the robustness of the findings, our study is not without limitations. First, due to budget and time constraints, we only surveyed two farming communities in the LRB, which may limit the generalizability of our monetary estimates to farmers in other areas of the Limpopo Water Management Area or even those of the entire LRB in South Africa. We are, however, certain that the farmers included in this study represent the population that would most likely be targeted for such a programme in the Limpopo Province. Second, as is common with all stated preference techniques, our monetary estimates are entirely limited to the study's design and the data collected. Our study can be further advanced with similar research in other parts of the Limpopo Water Management Area or even the entire LRB in South Africa.
In addition, to better inform policy decisions, it is essential to have a comprehensive cost–benefit analysis of the WTA compensation for controlling agNPS pollution under various policy scenarios. This analysis should consider how the costs of mitigation actions may fluctuate when the monetary values of agNPS pollution control attributes and the potential consequences of future policy alternatives are taken into account. By providing a more detailed understanding of the costs and benefits, policymakers can design more effective agNPS pollution control programmes that optimize resource allocation and achieve desired environmental outcomes.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S1355770X25100168.
Acknowledgements
This research was partially funded by the National Research Foundation (NRF) of the Republic of South Africa, under the project Investment Decisions in Water and Rural Development Programmes to Promote Food Security and Resilience of Smallholder Farmers in South Africa. The authors would like to thank the NRF for funding this research. The content of this work is solely the responsibility of the authors and does not necessarily represent the official views of the NRF. We also thank the enumerators and the survey respondents who devoted their time to make this research possible.
Competing interests
There are no competing interests to declare.
Open access
