Literature review on nonmonetary payments in choice experiments

As noted in the Introduction, the payment vehicle in most stated preference studies is typically presented in monetary terms, but the use of cash is low and is often supplemented with barter and work exchange in developing country settings (Rai and Scarborough Reference Rai and Scarborough2014; Gibson et al. Reference Gibson, Rigby, Polya and Russell2016; Vondolia and Navrud Reference Vondolia and Navrud2019). The use of monetary payments in such contexts may undervalue the resource or policy being valued (Abramson et al. Reference Abramson, Becker, Garb and Lazarovitch2011; Rai et al. Reference Rai, Shyamsundar, Nepal and Bhatta2015; Rai and Scarborough Reference Rai and Scarborough2014). In response to these concerns, there is a growing nonmarket valuation literature exploring the use of nonmonetary payment vehicles in the context of rural developing country settings. Some of the earliest work focused on using nonmonetary payment options (Swallow and Woudyalew Reference Swallow and Woudyalew1994; Shyamsundar and Kramer Reference Shyamsundar and Kramer1996; Echessah et al. Reference Echessah, Swallow, Kamara and Curry1997; Asquith et al. Reference Asquith, Vargas and Wunder2008), and more recent work compares outcomes like WTP and estimation efficiency across monetary and nonmonetary payments (Hassan et al. Reference Hassan, Olsen and Thorsen2018; Pondorfer and Rehdanz Reference Pondorfer and Rehdanz2018; Vondolia and Navrud Reference Vondolia and Navrud2019; Vondolia et al. Reference Vondolia, Eggert, Navrud and Stage2014; Abramson et al. Reference Abramson, Becker, Garb and Lazarovitch2011; Casiwan-Launio et al. Reference Casiwan-Launio, Shinbo and Morooka2011; Rai and Scarborough Reference Rai and Scarborough2013; Rai and Scarborough Reference Rai and Scarborough2014; Gibson et al. Reference Gibson, Rigby, Polya and Russell2016; Rai et al. Reference Rai, Shyamsundar, Nepal and Bhatta2015; Eom and Larson Reference Eom and Larson2006). We provide a discussion of some of these key studies and findings below.

Overall studies can be broadly separated into studies that use labor or time payments as the nonmonetary payment and other forms of payment as the nonmonetary payment. Within the literature that uses labor/time, the studies consist of split sample studies where respondent see either monetary or a nonmonetary payment choice card or studies that include combined choice cards with both nonmonetary and monetary choice cards (Hagedoorn et al. Reference Hagedoorn, Bubeck, Hudson, Brander, Pham and Lasage2021, 2020).

The following examples highlight the key results from the labor vs. monetary in split sample studies. Casiwan-Launio et al. (Reference Casiwan-Launio, Shinbo and Morooka2011) use a split sample contingent valuation study that considers money and labor hours as payment vehicles that is conducted with three villages in the Philippines. They monetize the labor payment using one-third the value of the wage rate and find that the converted WTP using labor is between 3 and 8 times larger than monetary WTP. Rai et al (Reference Rai, Shyamsundar, Nepal and Bhatta2015) study the WTP and preferences for watershed services in 12 village areas in central Nepal. The authors find that a larger proportion of the respondents are willing to pay using labor vs. a monetary payment (72% vs. 50%) and that the social benefits from environmental services are 1.4 to 2.2 times higher when using labor payments compared to monetary payments. Rai and Scarborough (Reference Rai and Scarborough2013, Reference Rai and Scarborough2014) study the WTP and preferences for management of invasive plants among rural communities in Nepal, and they include both monetary and labor payments. Rai and Scarborough (Reference Rai and Scarborough2014) allow the respondents to choose between monetary and labor payment options and find that 65% opted for the labor payments and 35% opted for the monetary payment. Hassan et al. (Reference Hassan, Olsen and Thorsen2018) compare three different payment vehicles in a study on wetland conservation in Malaysia: incomes taxes, voluntary donations, and a reduction in government subsidies for consumer goods. The authors conduct the survey with respondents living in six rural villages and four urban towns adjacent to the wetland. They find that reduction is subsidies as a payment vehicle has a higher price sensitivity and small unexplained variance and that this may be a favorable payment vehicle in terms of improved payment consequentiality. Navrud and Vondolia (Reference Navrud and Vondolia2020) study flood insurance among smallholder irrigation farmers in Ghana using a split sample with varying payment vehicles. They find that insurance up-take is lower when payments are required in labor compared to payments of money or payments using the rice harvest.

The existing literature finds mixed evidence about both the applicability and comparability of using nonmonetary payments vehicles in developing country valuation studies (Gibson et al. Reference Gibson, Rigby, Polya and Russell2016; Navrud and Vondolia Reference Navrud and Vondolia2020). We contribute to this growing literature by conducting a study that compares valuation results under monetary and labor payments vehicles in a rural developing country setting. In our study, we use a random assignment of respondents to labor and monetary payments eliminating the confounding issues that can occur when respondents selected the payment option. We also conduct the study in a new context focused on valuing the ecosystem services from restoring an ancient irrigation system known as cascade tank systems in Sri Lanka. We also contribute to the literature on conducting choice experiment studies in developing countries by exploring how results of choice experiments are impacted by conducting group information sessions.

Application – cascade tank systems

Cascading tank systems are an ancient irrigation system of connected reservoirs found in South Asia. The water storage system is designed to collect rainwater during the wet season to use in the dry season to support agriculture and population expansion to dry/arid areas (Panabokke Reference Panabokke2004; IUCN 2016; Dissanayake and Vidanage Reference Dissanayake and Vidanage2022). The cascading tank systems can be considered as an early application of Integrated Water Resources Management (IWRM) given the use of landscape-level planning and water and soil conservation aspects built into the landscape (Mendis Reference Mendis2002; IUCN 2016; Vidanage et al. Reference Vidanage, Perera and Kallesoe2005). The tanks (the term “tank” is derived from the Portuguese word “tanque” meaning a small man-made reservoir) are created by blocking the natural drainage system in a watershed using earth bunds in selected locations to enable storing of water and forming a series of tanks that allows the water to cascade from tank to tank (Panabokke Reference Panabokke2004; IUCN 2016).

The system, which has paralleled to the Japanese land use system of Satoyama, was established and fostered by ruling Kings. The small tanks are found in South India, China, Thailand, and Indonesia (Plan Sri Lanka 2012) with some differentiation. Brohier (Reference Brohier1934) in his monumental work on Ancient Irrigation Works referred to these clusters of tanks [cascades] in the following words:

These systems are hydrologically and socioeconomically interlinked (Sakthivadivel et al. Reference Sakthivadivel, Fernando, Panabokke and Wijeratne1996) and have evolved several millennia ago incorporating the principles of IWRM, and landscape approaches, and water and soil conservation aspects into planning and possibly also governance (Mendis Reference Mendis2002). Tennakoon states that small cascade-based tank systems had been a classic example of man’s ability to maintain a long-lasting symbolic relationship of man with available water, vegetation, climate, soil characteristics, animals domesticated, and wild being a partner of a well-renowned hydraulic civilization (Tennakoon Reference Tennakoon2001).

In Sri Lanka, there are estimated to be about 15,000–18,000 tanks and about 13,000 anicuts, diversions from streams (Panabokke Reference Panabokke2004; IUCN 2016). Recent works from the last few decades, led by Madduma Bandara (Reference Madduma Bandara, Lungquist and Halknmark1985), have re-discovered that these tanks are not occurring randomly but rather were created to be spatially organized based on the natural landscape to collect rainwater from well-defined microcatchments (Madduma Bandara Reference Madduma Bandara, Lungquist and Halknmark1985; Panabokke et al. Reference Panabokke, Sakthivadivel and Weerasinghe2002).

The individual tanks are components of large systems or units called “cascades,” defined as “a connected series of village irrigation tanks organized within a micro- (or meso-) catchment of the dry zone landscape, storing, conveying and utilizing water from an ephemeral rivulet” (Madduma Bandara Reference Madduma Bandara, Lungquist and Halknmark1985). Work by Dharmasena (Reference Dharmasena2004) on small tanks indicates that the village tank systems have been developed to cater to diverse micro- as well as macro-land uses by having different components such as gangoda, chena, welyaya, gasgommana, godawala, perahana, iswetiya, kattakaduwa, and kivul ela (Fig. 1).Footnote ³ The cascading systems have ecological features and cultural customs to reduce erosion and improve water quality. These include planting specific tree species as wind breaks, a funnel-shaped tank design to reduce evaporation, the creation of waterholes/fields for animals to reduce human–wildlife conflict, the use of phyto-remediation plants along the interceptor to improve water quality and remove impurities, the presence of desilting water holes and cultural practices to manage silt, the use of conservation buffers, and management as a common pool resources with shared labor (Madduma Bandara Reference Madduma Bandara2007; Dharmasena Reference Dharmasena2004).

Figure 1. Schematic diagram of a STCS. (Source: IUCN, 2016).

Even though many of the tanks were created over two millennia ago, they continue to provide irrigation for about 40% of the total irrigable area of the country, an astoundingly high percentage when considering that these were designed 2000–1500 years ago (IUCN 2016). Further, the tank systems provide multiple other ecosystem services beyond water for agriculture, including fisheries and livestock (Renwick Reference Renwick2001), flood prevention, control of soil erosion, improvement of water quality, water storage for use in the dry season (Schütt et al. Reference Schütt, Bebermeier, Meister and Withanachchi2013), climatic resiliency given changing climate patterns, human integrity, and retaining soil health (Senanayake et al. Reference Senanayake, Wijesekara, Hunter, Bélair, Ichikawa, Wong and Mulongoy2010).

At the same time due to various historical and sociocultural reasons, these systems were neglected over a long period of time and have slowly degraded due to lack of traditional governance and management structures, regular maintenance, and modern irrigation efforts that ignored the interconnectedness of the cascade system. With the growing understanding of these complex systems over the last few years, there is an increasing interest, nationally and globally, to better understand the cascade tank system, to restore the tanks to their original conditions, and to study the climate adaptation potential of the cascading tanks. Further, the restoration of degraded cascades has been identified as a key climate change adaptation mechanism in National Adaptation Plan for Sri Lanka (Ministry of Mahaweli Development and Environment [MMDE] 2016). The Food and Agricultural Organization of the FAO declared cascade tanks systems in Sri Lanka as a GIAHS in 2016 (FAO 2017). The HSBC-Global Water Program provided a $500,000 grant for the restoration of one cascade system in 2013. Most recently, UNDP and GCF co-funded a $52 million effort to restore 300 tanks to “strengthen the resilience of smallholder farmers, particularly women, in the Dry Zone through improved water management to enhance lives and livelihoods.”

Even though there is a growing interesting in restoring the cascading tanks and the surrounding ecosystems, there is a lack of knowledge about preferences and values of the stakeholders for the cascade restoration and scientifically determined impact evaluations of cascade restoration. We use a choice experiment survey to study the preferences for cascade restoration (and restoration of ancient irrigation systems in general). The main choice experiment was conducted in a cascade tank system in Anuradhapura, Sri Lanka, among rural villagers (Fig. 2).

Figure 2. Attributes and their levels as depicted in the survey instrument.

Survey design

Selecting attributes for the CE questionnaire

The key task in any CE exercise is the selection of the attributes and their levels (Kragt Reference Kragt2009). The attributes chosen to describe the change have to be relevant to both decision makers and the respondents of the questionnaire. For this study, the relevant attributes were identified through an extensive process that included a broad literature review, consultation with experts/scientists working on cascades, use of google maps, and focus group discussions of cascade-dependent stakeholders were utilized (for details, see Vidanage Reference Vidanage2018; Dissanayake and Vidanage Reference Dissanayake and Vidanage2022). The attributes to describe the outcomes of the cascade restoration process and the levels we identified through this process are presented in Table 1 and Fig. 3.

Table 1. Attributes and the levels

Figure 3. Sample choice card.

Experimental design

Discrete choice models describe respondents’ choices among a set of alternatives. In this research, the alternatives are the different cascade restoration options. One of the alternatives to choose from will be the present status of the cascade (Alternative A, the status quo or business as usual) for which the respondents do not have to pay any additional payment if they wish to select that as their choice. Alternatives will have different levels of the five attributes, designed as add-ons to Alternative A.

Levels of attributes for other two alternatives were generated in an experimental design. A design with four attributes with three levels and a six-level payment attribute will result in a full factorial design with 486 (3⁴ × 6) different combinations of attributes. Therefore, a manageable orthogonal fractional factorial design was created using the SAS statistical package following Kuhfeld (Reference Kuhfeld2010). The resulting experiment design was 100% D-efficient and generated 54 unique choice profiles. The 54 sets of choice profiles were blocked into 9 sets with 6 choice profiles in each set. As found in literature and to account for learning, the first and the second choice profiles in a set were repeated at the end as the seventh and eight choice profiles (Carlsson et al. Reference Carlsson, Mørkbak and Olsen2012). Therefore, each respondent answered eight sets of choice profiles/questions.Footnote ⁶ In the analysis, the first two choice responses were dropped to account for learning.

These attribute levels were illustrated graphically and used in producing printed color cards of choice profiles to be given to respondents to make their choices. A respondent only answered either cash or labor choice cards as determined in the survey design. In the survey, the nine sets of choice cards are repeated among respondents in a way that each set is equally distributed among the sample in both cash and labor versions where applicable. Sample choice set used in CE is given in Fig. 4.

Figure 4. Willingness to pay for the money treatment for the full sample.

*Note: Coefficient estimates for the main effects analysis for the money sample. The WTP values are provided per season. The daily wage rate is approximately LKR 1000.

The questionnaire comprised of four sections. The first section is about the general information on the survey, which included the confidentiality of the information collected, the consent of the respondent, and that the respondents can leave the survey at any point and finally, a detailed background to the issues to be addressed in the survey. Section two was devoted to introducing the attributes and their levels and graphical representation of them in the choice cards, and the third section was on the choice cards and getting choice responses. The final and the fourth section was on the respondent’s socioeconomic information.

Data collection

The tool for primary data collection of the research was a structured survey questionnaire, drafted with preliminary research undertaken and improved through expert and field consultations. The survey instrument was field-tested with the enumerators in the Kapiriggama Cascade, a cascade adjoining the sample cascade, the Pihimbiyagollewa cascade.

As the choice experiment surveys are somewhat complex, the enumerators were provided with detailed training. The training covered an introduction to the subject of STCS (enumerators were taken to the cascade as part of the training), training on the CEs, survey methods, and on field testing of the survey for them to get a better understanding before they undertook the survey. They were also given a chance to role-play by administering few surveys among themselves before they were exposed to field testing. In addition to the survey questionnaire, a set of color printed materials were provided to the enumerators as a survey aid. It included the materials for introducing the survey, complete set of labor, and cash choice cards (72 × 2 = 144) methodically arranged in hardcover ring binder folders. During the data collection, the enumerators went over each of the attributes with the respondents before the respondents answered the choice questions to ensure that the respondents were aware of the details of each of the attributes before answering the choice questions. Two images from the field work highlighting the individual surveys and the group surveys are presented in the Appendix (Figs. A1 and A2).

Field surveys

The primary data collection surveys were conducted with 10 final-year undergraduate students as enumerators from the faculty of Humanities, Rajarata University of Sri Lanka, supported by an experienced survey coordinator. A stratified random sample of 516 households was selected from residents of the Pihimbiyagollewa cascade covering six Grama Niladhari divisions in two Divisional Secretariat divisions of the Anuradhapura district.Footnote ⁷ The six Grama Niladhari divisions has a population of 5265; within the cascade itself, we sampled about 50% of the population. Stratification was done to capture variability in upstream, mid-catchment area, and the downstream of the cascade. The respondents were split with 60% (approximately 300) taking the survey in the group information setting and 40% (approximately 200) taking the survey individually.Footnote ⁸ The assignment of respondents/groups to the monetary vs. labor treatments was done randomly so that half of the respondents completed the monetary payment survey and half of the respondents completed the labor payment survey. Table 2 presents the summary statistics of full sample.Footnote ⁹

Table 2. Sociodemographic characteristics of the sample

Notes: Sample average income is Rs 69,284 per month. The average income for the population of the Anuradhapura district is Rs 64,400 per month in 2019 (DCS 2019). The administrative units that overlap the sample has a population of 5265 and gender distribution of 50.12% male and 49.87% female. Within the cascade system itself, we sampled 50% of the households (a population of roughly 1000).

Model and estimation

The analysis we use is based on random utility theory which lays the foundation for analyzing choice decisions (Lancaster Reference Lancaster1966; Hensher et al. Reference Hensher, Rose and Greene2015; Johnston et al. Reference Johnston, Boyle, Adamowicz, Bennett, Brouwer, Cameron, Hanemann, Hanley, Ryan, Scarpa and Tourangeau2017). In this theoretical framework, the utility derived by individual (i) from an alternative/choice profile (j) is a function of the attributes (x) presented in the choice cards, and unobservable factors can influence utility, which are captured by a random component (ε). Consistent with the random utility theory (Hensher et al. Reference Hensher, Rose and Greene2015), the random and unobservable terms are assumed to enter the utility function additively. Therefore, individual (i)’s utility (U) from choosing alternative (j) is expressed as follows:

(1) $${U_{ij}} = {x_{ij}} + {\varepsilon _{ij}} \ldots $$

In order to ensure internal consistency, a CE contains multiple-choice sets. Each choice set includes the status quo, representing no change in the prevailing levels of different attributes (x), and two (or more) alternatives scenarios, noting that each alternative identifies different levels of a number of attributes. Therefore, individual (i) selects alternative j over alternatives j′ when expected utility (U) is greater than expected utility from all other options (U). The probability (Pr) that individual i will choose alternative j over other alternatives j′ in a complete choice set R is given by:

(2) $${\rm{Pr}}(j\backslash R) = {\rm{Pr}}({U_{ij}} \gt {U_{ij}}^,\,{\rm{s}}{\rm{.t}}{\rm{.}}\forall j' \in {\rm{R}},{\rm{and}}\,j \ne j') \ldots $$

In order to identify the most preferred alternative, equation (2) can be econometrically estimated based on responses to a household or individual survey. Assuming that the error term is Identically, and Independently Distributed (IID) and indirect utility (V) is linear in attributes (x), then equation (2) can be estimated with a CL model (McFadden Reference McFadden and Zarembka1974).

CL model

CL model (McFadden Reference McFadden and Zarembka1974; Greene Reference Greene2012) assumes that the observable utility function would follow strictly additive form (Birol et al. Reference Birol, Karousakis and Koundouri2006). The model was specified so that the probability of selecting a particular cascade restoration and management scenario was a function of attributes of that scenario and the alternative-specific constant (ASC). The ASC captures the effect of unobservable factors on the selection of alternatives relative to the status quo. In this analysis, ASC is a dummy variable that is coded as 1 when either management scenario B or C was selected, and as 0 when neither management scenario option was selected. The CL model is expressed as:

(3) $${V_{ij}} = {\rm{ASC}} + {\beta _{ij}}{x_{ij}} \ldots $$

where V refers to indirect utility obtained by the i ^th individual for the j ^th alternative and β is the coefficient of the attributes (x) included in the experiment.

The CL model assumes the Independence of Irrelevant Alternatives (IIA) property, which states that the relative probability of the two options being chosen are unaffected by the introduction or removal of other alternative. If the IIA property is violated, then the CL will be biased, and hence, a discrete choice model that does not require IIA property such as MMNL model should be used instead.

MMNL model

A MMNL model, also known as a random parameter logit (RPL) model, can also be estimated by relaxing some of the constraints associated with the IIA assumption in the CL model. In addition to IIA, one other limiting assumption that CL assumes is that the homogeneous preference across respondents which usually does not hold. Therefore, in MMNL model both IIA and the preference homogeneity are relaxed. As per Greene (Reference Greene1997), accounting for IIA and preference heterogeneity enables MMNL to produce unbiased estimates of individual preferences of demand, participation, marginal, and total welfare. Furthermore, accounting for heterogeneity enables prescription of policies that take equity concerns into account. An understanding of who will be affected by a policy change in addition to understanding the aggregate economic value associated with such change is necessary (Boxall and Adamowicz Reference Boxall and Adamowicz2002). The RPL model (Train Reference Train1998), which accounts for unobserved, unconditional heterogeneity, should be used in order to account for the preference heterogeneity in a pure public good (Kontoleon Reference Kontoleon2003).

In the MMNL model, the observed component (βx) is decomposed into two parts: the sum of the population mean (Υ) and the individual deviation of the random parameter (η). In this model, socioeconomic variables (s) are introduced to detect sources of heterogeneity. Further, the interaction terms in γ identify the impacts of individual-specific characteristics on selected alternatives and the ASC. The MMNL model is expressed as:

(4) $${V_{ij}} = {\rm{ASC}} + \Upsilon {x_{ij}} + \eta {x_i} + \gamma {s_i} \ldots $$

Applications of the MMNL model have shown that this model is superior to the CL model in terms of overall fit and the welfare estimates (Kontoleon Reference Kontoleon2003; Morey and Greer Rossmann Reference Morey and Greer Rossmann2003). However, as Lancsar et al. (Reference Lancsar, Fiebig and Hole2017) highlighted, in choice modeling, it is natural to start with classical logit model. Hence, we present results for both the CL model and the MMNL model. The analysis was conducted using STATA and the clogit and the mixlogit functions (Hole Reference Hole2007).

Estimation equations used are provided below:

1. 1. For CL,

   (5) $${{\rm{U}}_j} = \mathop \sum \limits_{k = 1}^K {\beta _k}{x_{kj}} + {\beta _p}{p_j} + {\varepsilon _j}$$

   The marginal value of attribute k is the ratio between the parameter β_k and β_p:

   (6) $$MWTP = - {{{\beta _k}} \over {{\beta _p}}}$$

2. 2. For MMNL,

   (7) $$U_j^i = \sum\limits_{k = 1}^K {{\beta _{ki}}{x_{kj}}} + {\beta _{pi}}{p_{ij}} + {\varepsilon _{ij}}$$

The choice experiment data were coded using the level of each attribute (see table notes for Table 3 for details). Next, we present the results from the estimation and then discuss the policy implications. We analyze the data using standard CL and MMNL models (Hole Reference Hole2007; Hensher et al. Reference Hensher, Rose and Greene2015).