1 Introduction

Water scarcity is one of the biggest threats that can negatively affect not only economic development, but also environmental and human-health quality worldwide. These negative impacts are more severe for developing countries that face limited financial, technical, and regulatory capacities to allow mitigation of the water scarcity. Limited rainfall combined with mismanagement of available water resources and poor water supply infrastructures have led to disastrous situations that left many people—mainly the poor in the developing world—with limited access to water. Recent estimates by the World Health Organization (WHO) suggest that 2.1 billion people worldwide lack access to clean water services (WHO 2017).

Addressing water scarcity requires a combination of supply and demand management measures that not only promote technological change to support widespread adoption of water-efficient equipment, but also changes in behavior that contribute to water conservation. In the residential sector, examples of such technological changes could include adoption of low-flow showerheads, low-flow toilets, water-saving devices in taps, toilets and showers, greywater collection systems, and in-house greywater treatment technologies. Examples of behavioral changes could include turning off running water while brushing teeth, turning off the shower when soaping up, using dishwashers and washing machines when loads are full, and capturing water in buckets while showering.

Yet, no consensus has emerged when it comes to understanding the factors that drive adoption of water-saving technologies in urban households, especially in the developing world. The few examples that exist in the literature are focused on industrialized countries (Fuenfschilling and Truffer 2016; Ward et al. 2012)) and have mainly highlighted the institutional conditions and governance structures that support sustainable transition to water-saving technologies in the urban sector. In developing countries, most of the existing studies have examined the economic, social, and sanitary problems associated with the lack of access to water services (Fuente et al. 2016; Whittington et al., 2009; Banerjee and Morella 2011; Jeuland et al. 2011; Cook et al. 2016), as well as the implications associated with a change in water tariff on access to water by poor households (Nauges and Whittington 2010; Banerjee et al. 2010; Whittington et al. 2015; Wang et al. 2005).Footnote 1

A number of policy instruments have been suggested for the residential sector to address water scarcity. Instruments range from applying sound water pricing that accounts for scarcity of the resource (Dinar et al. 2015; Diakité et al. 2009; Howitt et al. 2002; Howe 2007; Huang et al. 2010) to implementation water restriction programs (Kenney et al. 2004; Howe and Goemans 2002) that limit the authorized volume of water use per day. Additional instruments emphasize the need to raise awareness through education and capacity building to make users more cautious about the economic, sanitary, and environmental consequences associated with water scarcity. Another type of non-price demand management approach that has not been fully investigated is adoption of water-saving equipment within residential households. The very few examples encountered in the literature are from developed countries, where public authorities provide incentive schemes to encourage adoption of water-saving equipment. Such incentives include low-flow toilet rebate programs, and free distribution of plumbing retrofit kits offered by the local water agencies in California to mitigate the various droughts encountered in the region (LADWP 2015). Other examples are current rebate programs offered to households, in Canada and Australia, willing to adopt water-saving equipment (Statistics Canada 2009; Australian Bureau of Statistics 2006). We are not aware of any study that investigates the factors that drive adoption of water-saving equipment by households in developing countries. This is despite the fact that some developing countries experience severe water scarcity.

The purpose of this paper is, firstly, to investigate the factors that drive adoption of water-saving equipment within urban households in the city Cape Town, South Africa. Secondly, we analyse the preferences and scale heterogeneity of potential adopters by making use of the random utility theory (RUT) framework. The city recently experienced one of its worst drought seasonsFootnote 2 in four decades. Limited rainfall challenged water allocation decisions and placed many households in difficult situations that required a strict reshaping of water consumption behaviors.

We construct a theoretical model that builds on disaggregated technology diffusion framework, which identifies the diffusion path of a water-saving technology within households. The model is tested using choice modeling, where results from latent class analysis (LCA) are compared with estimates from both multinomial and conditional logit models. The results of our analysis show that key technological attributes (lifespan of the technology and its ease of use by the purchaser, the bad smell and likelihood to generate waterborne diseases) can influence household adoption decisions. Beyond its academic contribution, our work offers policy relevance in guiding public policy decisions that attempt to improve water use efficiency within the residential sector in urban settings in the developing world. Despite the fact that Cape Town is a relatively well developed city, yet with a high level of inequality,Footnote 3 this work can provide insights to many cities that experience water scarcity. We show that a combination of technological and targeted socioeconomic policies that support education and information dissemination about the features of the selected technologies can help enhance the adoption of greywater technology in urban areas.

The paper is structured as follows: Section Two provides an exhaustive review of the existing literature on adoption of efficient water use technologies in the urban sector. The methodology of this paper is presented in Section Three. Section Four describes the different stages undertaken to design the choice of experiments and the data used. The results and discussions are presented in Section Five. Section Six concludes the paper and provides key policy implications.

2 Related literature

The importance of public policy in fostering adoption of more efficient technologies in various water consuming sectors has been widely studied in the economic literature (Katz and Shapiro 1986; Nelson 1982; Koundouri et al. 2006; Dinar and Zilberman 1991; Hall and Khan 2003; Dosi 1982; Edquist 2004; Kerr and Newell 2003; Millock et al. 2012). Existing studies range from the energy (Jaffe and Stavins 1994; Jaffe et al. 2003; Li and Just 2018; Haq and Weiss 2018; Versteeg et al. 2017), to the agricultural (Koundouri et al. 2006; Zilberman et al. 2013; Emerick et al. 2016; Sunding and Zilberman. 2001; Schoengold and Zilberman. 2007; Winters et al. 2004; Bontemps and Couture. 2002; Honlonkou. 2004; Speelman et al. 2010) and health sectors (Hyysalo 2010; Drummond et al. 2013; Faulkner 2009; Buxton and Chambers 2011; Dishman 2012).

Usually, analysis of technology diffusion highlights two major factors as main drivers of the adoption of more efficient, superior technologies: technology-push and demand-pull (Nelson 1982; Edler and Yeow 2016; Ghisetti 2017; Dosi 1982). Demand-pull assumes that technology diffusion remains mainly driven by the demand that emanates from consumers. Producers innovate and create a market that supplies those technologies, which match consumers’ demands and trigger a technology push.

It is important to highlight that such clear-cut distinction between demand-pull and technology-push hardly reflects the reality of technology diffusion. This has been highlighted in Roger (2003) as well as in Hall and Khan (2003), who argued that many technologies are disseminated, because the right combination of market, governmental and institutional policies have been provided, which create incentives for consumers and producers to adopt such new technologies. With regard to water-conserving technology in urban settings, as indicated earlier, the literature on the factors that drive adoption of water-saving equipment in developing countries is very limited. Most of the existing studies are focused on the developed countries, where access to data is much easier.

For instance, Renwick and Archibald (1998) developed a theoretical framework that helps us understand the extent to which water demand side management (DSM) policies might affect residential demand for different classes in Southern California. The theoretical model is tested using a two-stage least square (2SLS) estimation procedure in a natural experiment, and data collected from 119 single family dwellings in Santa Barbara and Goleta, California. The findings show that adoption of water-efficient garden irrigation technologies has a positive and significant effect on reducing water consumption. The authors argue that policies which aim at promoting water conservation measures positively influence a probability of the adoption of water-efficient equipment. Campbell et al. (2004) investigate the impacts of various policies (both price and non-price water demand side management) on promoting water-saving behaviors and adoption of water-saving equipment in Arizona. Data collected from 19,000 households over 6 years was used in a multivariate regression analysis. The results show that even a small increase in water pricing leads to adoption of water conservation measures. Their findings confirm that even an imposition of non-price policy (rules, increased awareness, and providing engineering technologies) may equally lead to a decrease in water consumption and adoption of more water-efficient equipment. Geller, Erickson and Buttram (1983) show the positive and significant impact of non-price DSM on reduction in water consumption. The results highlight the importance of supporting educational, behavioral, and technological changes within households for promoting water-use efficiency. Nauges and Whittington (2010) provide a comprehensive overview of the factors that influence water demand in the developing world.

The study highlights the need for a better understanding of the characteristics of water consumers in these underdeveloped countries. The improved understanding is expected to encourage a better design and implementation of efficient water conservation measures. Three classifications of water consumers are identified in their paper: households with incomes ranging between US $150 and $400 per month that can afford municipal piped water services; households in slums with income levels below US $150 per month; and households in the rural areas in Sub-Saharan Africa and South East Asia with income levels less than US $30 per month.

The authors highlight the need to find appropriate and targeted policy mechanisms that not only foster an increase in water access but also favor sustainable water consumption. Using data collected from urban households in the city of Granada in Spain, Perez-Urdiales et al. (2016) evaluate the extent to which water tariff structures influence water consumption. Their findings show the need to carefully choose the type of instruments used to induce reduction in water consumption. The authors argue that a combination of price and non-price mechanisms might help reduce water consumption and promote water use efficiency. Millock et al. (2012) develop a theoretical framework that studies the impacts of non-price tax policy on adoption of monitoring technology to control stock of externality. Although their study was not referring specifically to household water consumption, the underlying theoretical framework provides good insight that helps quantify the factors that drive adoption of environmentally sound technologies. Renwick and Green (2000) evaluate the impacts of various DSMs on urban water resource management in eight water agencies in California, serving 7.1 million people. Their targeted policy instruments were water allocation, use restrictions, and public education. The results suggest that targeted policies were effective in inducing water use efficiency in the sample areas.

The magnitude of these changes may be different depending on the seasons in which the households experience policy interventions. For instance, responses to a price change was 25% higher in summer, reflecting outdoor water consumption associated with high temperature. Grafton et al. (2011) show the extent to which implementation of various pricing schemes may affect water consumption in ten OECD countries. Their results show that price and non-price mechanisms influence household adoption of water-saving technologies and change in behaviors. Millock and Nauges (2010) investigate the factors that drive adoption of four different types of water-saving technologies: waterwise washing machines, low-volume flush toilets, restrictor taps in water supply, and rainwater collector tanks in ten OECD countries. Their results show that adoption of water-saving equipment remains strongly driven by key factors, such as household size, ownership of the property, water pricing, as well as degree of sensitivity towards environmental values. Olmstead and Stavins (2009) compare price and non-price approaches to urban water conservation. Their paper shows the relative merits of market-based and prescriptive approaches to water conservation. The results show that using price to manage water demand is more cost effective than implementing non-price conservation measures.

Their paper highlights the importance of including key important factors (equity and distributional consideration, political consideration, and the costs of monitoring and enforcement) in designing any water-demand management options, especially when it comes to adoption of water-saving technologies. Table 1 provides a summary of key previous findings that investigated the impacts of supply and demand management policies on promoting adoption of water-saving technologies. Such studies took place mostly in industrialized countries.

Table 1 Choice of water conservation measures: review of previous literature
Full size table

Our paper contributes to the current discussion on factors that drive adoption of water-saving technologies using empirical evidence in the form of a natural experiment from a developing country. We elicit preferences to inform water management options that aim at facilitating widespread adoption of water-saving equipment. Rapid population growth and rural-to-urban migration observed in the developing countries, combined with poor quality of water supply infrastructures and weak regulatory capacity have led to situations in which urban dwellers may find it more appropriate to respond to water-demand management measures that reduce volume of water consumption within the household, especially through adoption of water-saving equipment.

The city of Cape Town faced a water crisis situation that was brought on by three consecutive years of insignificant rainfall. In January 2018, Cape Town city officials announced that the reservoir serving four million people, was 3 months away from running out for municipal water consumption. That water crisis was coined “Day Zero.” City inhabitants were requested to drastically cut their water consumption to reduce the risk of having no running water from their taps. Was that a sufficient “threat” to induce changes in the behavior of household water consumption, including adoption of water-conserving technologies?

The underlying hypothesis of this paper is that adoption of water-saving technologies may help households reduce the volume of water consumption, allowing the city of Cape Town to meet its water budget. The selected water-saving method referred to throughout the paper is greywater recycling technology. Greywater usually consists of wastewater coming from baths, showers, kitchen sinks, and washing machines. It contains lower concentrations of microbial contents and chemical characteristics compared with sewage water (Roesner et al. 2006). Previous studies show that usage and treatment of greywater by households and communities allow not only a reduced demand for freshwater, but it also saves in public expenditures by centralized wastewater treatment plants (Gross et al. 2005; Jefferson et al. 1999; Wiltshire 2005; Morel and Diener 2006). This paper targets two widely used greywater technologies: the stand-alone greywater tank (technology 1), and a sophisticated and integrated system of greywater that is connected to the toilet and can be used to flush the toilet instead of using potable water (technology 2). Both technologies are used to collect, store, and treat greywater from various sources within the household (bath, kitchen, washing machines, showers, and so forth.). The difference between the two is as follows: technology 1 is not connected to the plumbing structure that is linked to the toilet, whereas technology 2 is. Therefore, with technology 1, the collected greywater is treated and transported to the end-use point, whereas technology 2 only treats the greywater and returns it to the system. Treatment consists of a combination of anaerobic and aerobic processes with disinfection options that eliminate aesthetic, health, and other problems that are caused by organic matter, pathogens, and solids, and they meet reuse standards. These technologies are currently being manufactured in Cape Town, South Africa.

3 Methodology

3.1 Theoretical model of diffusion of water-saving technology

The proposed methodological framework builds on the disaggregate technology diffusion literature to account for consumer heterogeneity impact on adoption of a new technology. We introduce a simple theoretical model that explains the diffusion process of water-saving technology within urban households. Our theoretical approach builds on Bass (1969) but extends it by introducing assumptions that allow us to capture the characteristics of different types of urban households. Let’s assume that adoption of greywater equipment within a household is a function of innovation \((p>0)\) and imitation \((q>0)\). Equation (1) shows the probability of diffusion \({P}_{t}\) of the new technology, when \(p\) and \(q\) are exogenous:

$${P}_{t}=P-\frac{q}{M}B\left(\dots \right)$$
(1)

where p and q represent the parameters that capture the desire of certain individuals to innovate and to imitate, respectively. Innovation refers to the desire that some households have to install and experience (innovate) new water-efficient technologies in their homes. Imitation refers to those households that have installed such technology only after having observed that their neighbors have previously adopted that technology. The latter type of household is supposed to imitate the former one. Arrow (1962) refers to this classification as the first mover and the follower, respectively. \(M\) is a parameter that captures the potential market share of the technology and \(B (\dots )\) is the cumulative number of households that are willing to adopt the technology. If we define \(f\left(t\right)\) as the likelihood of purchase at time t, with \(F\left(t\right)={\int }_{t=0}^{T}f(\tau )d\tau \) and \(F\left(0\right)=0\), Eq. (1), the probability of adoption, can be transformed into Eq. (2), following Srinivasan and Mason (1986):

$${P}_{t}=\frac{f(t)}{1-F(t)}=p+\frac{q}{M}F(t)$$
(2)

In the early Bass (1969) model, the cumulative function B(…) has been considered linear, despite the fact that diffusion of a new technology hardly takes place in an environment that is stable, linear, and unchanging. Mansfield (1985) for instance, argued that diffusion of a technology is driven by a combination of factors, such as risk profile of the adopters, their level of income, and the institutional settings within which they operate. We follow Mansfield (1985) and argue that elements such as risk profile might also affect the extent to which a technology gets diffused within a society. When a new innovation is introduced, it has been shown that rate of adoption might differ between not only rich and poor households (Khan and Ravikumar 2002) but also between potential adopters with different risk profiles (Foster and Rosenzweig 2010). We account for the risk profile of the potential adopters and assume a cumulative function that includes a parameter \(\mu \) that captures the risk profile. Individuals are risk-neutral when \(\left(\mu =1\right)\), and risk-averse when \(\left(\mu =0\right)\). Therefore, assuming that \(B\left( \right)=M\left(1+\frac{1}{\mu }\right)F(t)\) and taking into account Eq. (2), we determine the likelihood of purchase of greywater technology at time t. This likelihood is represented in Eq. (3):

$$f\left(t\right)=\left[p+q\left(1+\frac{1}{\mu }\right)F(t)\right]\left[1-F(t)\right]$$
(3)

Finally, we assume the existence of a state variable \(s\left(t\right)=Mf(t)\) that captures the evolution of sales of the new technology. Transformation of Eqs. (1–3) leads to Eqs. (4) and (5):

$${P}_{t}=\frac{s(t)}{B(t)\left[M-\frac{\mu }{\mu +1}\right]}=p+\frac{q}{M}B(t)$$
(4)
$$s\left(t\right)=p\left(M-\frac{\mu }{\mu +1}\right)B\left(t\right)+\frac{1}{1+\mu }q{B}^{2}\left(t\right)$$
(5)

The intuition behind (4) and (5) is straightforward. The evolution of sales (or the diffusion) of the water-saving technology is a function of the risk profile, the innovation and imitation coefficients and the cumulative function, B.

3.2 Empirical strategy

We make use of choice modelling to validate the theoretical approach developed above. Our decision variable is whether a household decides to adopt water-saving equipment or not. This is a binary decision which can be analyzed using the discrete choice experiment (DCE). The random utility theory (RUT) is referred to design of DCE (Lancaster 1966; Mansky 1977; McFadden 1973), since it offers a good theoretical ground to study choice behaviors. It originated from the concept of utility maximization principle (McFadden 1973) and represents a suitable approach in our present paper, since it allows us to identify the attributes and socio-economic characteristics that drive adoption of water-saving equipment by urban households in South Africa. This framework is based on probabilistic choice theory, where households are expected to choose alternatives that provide them with the highest utility (utility maximization approach). The utilities are driven by choices made by households and those choices are based on the attributes of the selected water-saving technologies. Households derive satisfaction from the attributes that the water-saving technologies possess. The attributes have been given different levels, which allows us to draw different combinations that form part of the choice sets presented to respondents. According to the RUT, households choose their preferred alternatives with the highest expected utility. A change in an attribute (or its level) is, therefore, expected to affect the choice probability of choosing a given water-saving technology. For any sampled household i (i = 1….I), the utility \(\left({U}_{ij}\right)\) derived from a given alternative (j) associated with the selection of water-saving equipment, is given by Eq. (6):

$${U}_{ij}={\beta }_{i}{x}_{ij}+{\varepsilon }_{ij}$$
(6)

where \({\beta }_{i}\) is a vector of individual-specific coefficients, \({x}_{ij}\) is a vector of observed attributes relating to individual i and alternative j, and \({\varepsilon }_{ij}\) is a random term. Therefore, the indirect utility is made of a deterministic component denoted by \({V}_{ij}={\beta }_{i}{x}_{ij}\) and a random term (Mansky 1977; Louviere et al 2000). The researcher can estimate the deterministic component, whereas the random component captures the difference between \({U}_{ij}\) and \({V}_{ij}\). The probability that households i choose alternative k over another one \(j\) from a set \(M\) is given by Eq. (7) below:

$${P}_{ik}=P\left({U}_{ik}>{U}_{ij}\right)=P\left({\varepsilon }_{ik}-{\varepsilon }_{ij}<{\beta }_{i}{x}_{ik}-{\beta }_{ij}\right) \forall k\ne j$$
(7)

Based on the random utility framework, different models are estimated under varying assumptions on preference and scale heterogeneity. We first estimated the multinomial logit model (MNL) as the baseline model. MNL assumes that each unobserved factor is independently, and identically distributed (iid) extreme value and preferences across households remain homogenous (Hess and Rose 2009). The model provides an estimation of the direction of the impact of an attribute (with the reference to the base) in explaining a household’s probability of choosing an alternative scenario. MNL is a useful model when one would like to investigate the contribution of a given attribute to influencing household choice of water-saving technology. However, the MNL model has several weaknesses.Footnote 4 We, therefore, proceeded to estimate the conditional logit (CL) model which relaxes some of the assumptions of the MNL model. It is important to note that the CL model relies on the IIA assumption,Footnote 5 which states that the probability of choosing between different options is not affected by other alternatives.

We employed the Hausman and McFadden tests under the null hypothesis of violation to test the IIA assumption (Hausman and McFadden 1984). The results reveal no violation of the IIA assumption. Although heterogeneity can be introduced in the CL model, it does not account for class heterogeneity. Latent class analysis (LCA) is then used to account for class heterogeneity in addition to preference heterogeneity and to release the restrictive assumption of IID of error terms. For this reason, the LCA is the main methodology of this paper. The underlying argument supported by economic stylized facts, is that consumers who belong to the same class (such as education, income) tend to have the same behavioral patterns (Swait 1994; Hess and Rose 2009; Hess 2014; Gupta and Chintagunta 1994). LCA assumes that a discrete number of classes are sufficient to account for preference heterogeneity across classes (Hess 2014). Therefore, the unobserved heterogeneity is captured by these latent classes in the population, each of which is associated with a different parameter vector in the corresponding utility.

Therefore, in our model of S classes \({\beta }_{i}(L=l,\dots .,s)\) specific parameters would be estimated with the possibility of some \({\beta }_{i}\) remaining constant across some of the classes. The technique is also used to account for attribute non-attendance (ANA), in which responders ignore one or more of the attributes when making their choices. LCA accounts for preference heterogeneity by simultaneously estimating probability of class membership and preference parameters, based on individual characteristics. The technique uses a probabilistic class allocation model.

Based on AIC and BIC criteria, three different behavioral classes were identified: (i) risk averse, (ii) innovators, and (iii) supporters of greywater technology. The class “risk-averse” includes responders who believe that adoption of greywater treatment technologies leads to health-related problems. By definition, risk-averse households are those, who, when exposed to uncertainty, attempt to lower that uncertainty (Dixit and Pindyck 1994). The class “innovator” refers to households that have already (at the time of the survey) adopted greywater technology. Finally, the last class, “supporters of greywater technology,” is made up of households who think that in-house treatment of greywater must be allowed through implementation of specific legislation. Households within that group have an inherent belief system that supports dissemination of greywater technologies.

Following the above classification, we use the RUT discussed above to determine the socioeconomic and attributes that influence adoption of water-saving practices. This is done by assuming that individual i belonging to class s has a utility \({U}_{is}\) that is derived from adoption (j = 1) or non-adoption (j = 0) of greywater treatment technologies. Examples of positive utility provided by adoption of such technologies are reduction in global water consumption, reduced water bills, and lower reliance on water supplied by the municipality for non-potable water consumption within the households. This formulation is a typical representation of a random utility model (RUM) that has deterministic and random components (Lancaster 1966) for each class. The functional form of the utility function is given in Eqs. (8) and (9):

$${U}_{in/s}={V}_{in/s}+{\varepsilon }_{in/s}$$
(8)

with

$${V}_{inj/s}=f\left({X}_{inj};{Z}_{it}/{\beta }_{s}\right)$$
(9)

where V, represented by \(f\left({X}_{inj};{\beta }_{s};{Z}_{it}\right)\), is a function of the water-saving technology attributes of alternative \(n\) faced by individual \(i\); \({X}_{inj}\) is the matrix of attribute levels for the new greywater equipment; \({Z}_{it}\) the vector of individual characteristics, and \({\beta }_{s}\), the vector parameters specific to the selected class. The probability of a choice is given by Eq. (10) below:

$${P}_{i}^{\mathrm{LCA}}=\sum_{s=1}^{S}{P}_{is}\prod_{t=1}^{T}\prod_{j=1}^{J}{\left\{\frac{\mathrm{exp}\left({x}_{ijt}^{^{\prime}}\beta \right)}{\sum_{j=1}^{J}\mathrm{exp}\left({x}_{ijt}^{^{\prime}}\beta \right)}\right\}}^{yijt}$$
(10)

where Pis is the probability of household i belonging to class s. The log-likelihood for this model is \(LL=\sum_{i=1}^{I}ln\left({P}_{i}^{\mathrm{LCA}}\right)\). Table 2 shows the number of classes determined. Finally, we estimate the marginal effects of each attribute to allow our results to be more policy relevant. Equation (11) represents the willingness to pay (WTP) for the greywater technology:

Table 2 Number of classes
Full size table
$$\mathrm{WTP}=-\left(\frac{{\beta }_{s}\mathrm{attribute}}{{\beta }_{s}\mathrm{cost}}\right)$$
(11)

4 Design of the choice experiment

4.1 Design of the attribute space

The selection of attributes included in the survey has important implications for the results of the choice experiment. We rely on the previous literature to identify the factors that drive adoption of new technologies within households. Table 3 provides an overview of the attributes and the attribute levels considered. Previous studies highlight the lifetime of the technology as one of the major factors that influence adoption of technology (Comin and Habijn 2010; Ahsanuzzaman 2015). Households are more willing to adopt new equipment that is perceived to be reliable and has a longer lifetime. The two levels of this attribute are: short and long term. The second attribute identified was ease of use, which is perceived as an important determinant in technology adoption. For quite a number of technologies, diffusion has not been widely observed, because potential adopters find it costly and time-consuming to acquire skills that are needed to use the technology (Mukoyama 2004; Bartel and Lichtenberg 1987; Doms et al. 1997). Beyond learning cost and time spent at the early stage of adoption, ease of use encompasses any effort made to run, repair, and maintain the selected technology at its best standard. Two levels are selected for the second attribute: easy (when no extra training is required before usage of the technology), and difficult (when intensive pre- and post-usage training is necessary for a well-functioning technology).

Table 3 Definition of attributes and their levels
Full size table

Technology adoption is also driven by externalities associated with the usage of that technology. After consultation with water policymakers in the region, we include two plausible externality effects: smell, and health-related externalities. Some greywater treatment technologies release a smell, which affects not only the household in which the technology is installed, but also the neighboring households. Many different cases are observed in which neighbors complain about the bad smell coming out of greywater systems installed in their surroundings. Two levels are assigned to each of these attributes: low and high (for smell) and high-risk and safe (for health-related externalities), respectively. Health-related externalities (waterborne diseases, trauma, and discomfort) can result from a bad smell or any physical contact made with the greywater without having taken precautionary sanitary measures.

Several studies have shown the extent to which health-related benefits drive adoption of some technologies (World Health Organization (WHO), 2017; Drummond et al. 2013; Faulkner 2009; Buxton and Chambers 2011; Dishman 2012). Finally, costs were introduced as an attribute to account for the technology. In addition to our literature review we conducted several focus group discussions (FGDs) with water policymakers, private companies that design greywater treatment technologies, and selected households.

Participants of the FGDs were recruited through their local associations or non-government organizations (NGOs). We conducted a series of workshops with policymakers and local authorities who assisted us in recruiting the participants. A total of four FGDs were conducted in different locations, making sure that participants came from different segments of the society, namely, low-, middle-, and high-income groups. On average each FGD had 12 participants in total and four participants were invited from each segment. Through FGDs, we were able to develop a local understanding of important concepts associated with attributes and a way to convey them to the respondents. All the identified attributes were discussed with these stakeholders, and they were validated by the experts. The expert group was composed of researchers and officials from the municipality in Cape Town.

4.2 Experimental design

The main objective of experimental design in choice experiments is to develop designs that yield efficient and unbiased estimates of preference parameters and value estimates (Johnston et al. 2017). Table 4 provides an example of the choice experiment scenarios that were presented to the respondents. With four attributes varying across two levels each, and one attribute varying across five levels, there were 80 (24 × 51) possible combinations of the attributes and their levels. To minimize bias, a full factorial orthogonal design of 24 alternative profiles was created using NGENEFootnote 6 from the full set of possible combinations. The number of alternatives is informed by the literature review and is based on the frequently used number of blocks and choice sets for a design similar to the one being considered in this paper. We leave the intermediate levels out to make sure there is a clear trade-off for the respondents when it comes to making choices.

Table 4 Sample of choice set
Full size table

The software produced an efficient design with one status quo and two non-status quo alternatives per choice set, and four choice sets arranged in six survey blocks/cards. Each choice set presents the characteristics of the technologies. Respondents were randomly assigned one of the six survey blocks/cards which had been prepopulated in six different questionnaire versions. The status quo represents the technology that the household is currently using and is known by the responder. This could constitute a mix of the different attributes or even the absence of a water-saving technology. Therefore, it is inappropriate to specify the attributes for the status quo since these can vary from household to household. All the attributes except the cost ones, were dummy coded.

4.3 Survey design

Stated preference (SP) studies should elicit evidence that pieces of information are understood, accepted, and viewed as credible by respondents (Johnston et al. 2017). To elicit household preferences for attributes, the study used a survey based choice experiment. The survey method allowed enumerators to convey the message and explain difficult concepts. The questionnaire had five sections. Section one collected information about the demographic characteristics of the respondents; section two collected information on household income and expenditures; section three collected information related to water-reuse greywater technology; section four collected information on nature inclination of the respondents, and section five ran the choice experiment with the different types of water-saving equipment.

The respondents were told that the city of Cape Town is planning to introduce water conservation measures in response to water shortages as a result of frequent droughts. As part of the water conservation measure, city authorities will disseminate information about various water-saving technologies available in the market and supply them to buyers at a cost. Since dissemination of technologies has other cost implications for both the supplier and the buyers themselves such as information cost, the city will bear most the burden associated with the cost of getting information to cushion its customers. This cost is accounted for by the attributes. With respect to this, the respondents were asked to pay a once off levy to the city in exchange for a water-saving technology of their own choice.Footnote 7 The respondents could choose to maintain the status quo and pay nothing. As already alluded to, the status quo represents the current situation of the household, i.e., what they are doing now, whether they have a technology installed or not. If the responders choose the status quo, then there is no change in utility. As a result, the choice for ‘not purchase’ is captured in the status quo. The payment vehicle selection was informed by pre-testing to minimize unintended effects on value estimates (Johnston et al. 2017).

Respondents were then presented with a series of choice alternatives, differing in terms of their attributes and levels, and asked to choose their most “preferred water-saving technology” among a range of alternatives presented to them. Five undergraduate students were recruited from the University of Cape Town as enumerators, and a 2 day training session was administered in order for them to internalize the information being conveyed by the choice sets. As recommended by Johnston et al. (2017), a pilot study was also conducted on the third day in an area not selected for the main survey as part of training, to ensure that the respondents understood the attributes and to refine the survey instrument.

To facilitate the interview, we provided each respondent with a separate fact card describing the attributes in English. Each option in the choice set provided respondents with different attributes of a technology (e.g., a technology with a short lifetime, easy to use, with high likelihood of generating a bad smell or disease). Due to the subjective nature of verbal description and to ensure that respondents have a common understanding of the subject matter, each technology was visualized through digital manipulation of a control picture (the images of the technologies are included in Table 4). This was made to ensure that different types of technologies and changes in the attribute levels were easily illustrated without biases that may have arisen from differences in the respondents’ levels of education.

4.4 Data

The data was collected in Cape Town, South Africa. Various districts of the city were sampled to account for various important characteristics of the city and to achieve representativeness.

We collected data from 465 households in April 2018.Footnote 8 A team of five enumerators was recruited and spread around the city. Areas from the Southern (Plumstead, Rondebosch, Kensington, Wynberg, Southfield), Central (Goodwood, Maitland) and the Northern (Brackenfell, Northpine, Oakdale) parts of Cape Town were randomly selected. We obtained information about the households and streets in each suburb from the city authorities. Each day, a street name was randomly selected and allocated to a designated enumerator. Thereafter, a systematic random sampling was applied by the enumerator to identify and select the next respondent by skipping households according to a sampling interval (n), computed as the total number of households in a given area, divided by the sample size in that area. For instance, when the number of questionnaires required in a specific area were not received, enumerators were instructed to make a left turn before they reached the end of the street and continue.

Alternatively, we instructed enumerators to request a different street name to be randomly generated from the office. In addition to the choice experiment, the questionnaire collected information on household socio-economic variables, demographic characteristics, residential water use patterns, water consumption behavior, adoption and use of water saving technologies. To elicit information from greywater supporters the following questions were asked: “Should greywater use or treatment be allowed in homes?” “Do you think greywater is harmful to human beings?” “Does the use of greywater technology help to reduce your water bills?” We used factor analysis to recover an index for greywater supporters. To determine risk aversion, the following question was asked: “Do you believe that adoption of greywater treatment technologies leads to health-related problems?” To determine innovators, we asked: “Has your household already adopted greywater treatment technologies?”.

Figure 1 in the appendix presents the areas that have been surveyed. Table 5 shows the summary statistics of the selected variables to be included in the analysis. The selected variables are school years that capture the number of years spent in school by the household head, the gender of the household head; Day zero that represents a situation where the city of Cape Town runs out of water; the three classes: innovators, risk averse and supporters of greywater technologies; the water bills paid by the households; household income, cost of the technology and the alternative specific constant (ASC). The data was captured such that each individual was entered 12 times to include the choices they made for three options and four different choice sets. Responders averaged 53 years of age, and their average schooling was 13 years. A greater proportion (62%) of the sample was composed of male respondents. Quite a significant proportion (46%) of the respondents did not believe that Day Zero actually exists.

Table 5 Summary statistics
Full size table

The average income of the household sample is R17,048.83 (US $1,183) per month, while the average water bill is R159.30 per month. The average cost of a technology is R9,732.30, based on market prices collected at the time of the survey. We observed a great variability associated with income and technology cost, suggesting that most of the responders were middle aged, educated up to matric level, and belonged in the middle income category. The alternative specific constant (ASC) in Table 5 commonly reflects the status quo bias, and it measures the difference between status quo and non-status quo alternatives. Apart from sampling errors, other types of errors could have occurred and hence there was need to manage the data collection process so as to minimize these errors. Measurement errors, processing errors, non-response errors which occur when the concept implied by the survey question differs from the concept meant to be measured in the survey. To minimize these errors, enumerators were trained for several days so that they could familiarize themselves with the questionnaires before going to the field.

5 Results and discussion

We start by analyzing the results of the CL and MNL together, and we then proceed to discuss the LCA results. Table 6 represents the results of the conditional logit (CL) and the multinomial logit (MNL) models with and without interaction terms. The models with covariate interactions were run to evaluate the effects of individual characteristics on technology preferences. Table 7 presents the results of the LCA model.

Table 6 Results of the MNL and CL (with and without interaction)
Full size table
Table 7 Latent class analysis (LCA) results
Full size table

5.1 CL, MNL, MNL-interaction and CL-interaction results

The specification of the CL and MNL uses the attribute levels and the ASC to explain the alternatives selected by responders in the choice sets (Vermunt et al. 2008). The coefficients in the CL and MNL models are all statistically significant at the 1% level. The signs of the attributes lifetime and disease are consistent with expectations, whereas the sign of bad smell despite being significant is not consistent with expectation. We expect households to consider bad smell as a negative externality, which decreases the probability of adoption. The positive and significant coefficient of bad smell might suggest that the adoption probability is low if the technologies do not generate bad smell, implying that responders may prefer water-saving technologies that generate bad smell. This is because households may still choose greywater-saving technologies although they generate a bad smell during their use, as long as the technologies contribute to saving water and thereby reducing the water bill. Despite their ability to generate bad smell, greywater technologies still remain an appealing water-saving strategy in the study areas. This is supported by the fact that adoption of greywater can be combined with utilization of chemical scrubbers (gas and liquid oxidation, carbon/permanganate absorption, FeCI3 addition, for example) that allow a mitigation of the negative effects associated with bad smell to reduce the odour sources within the household.

Similar observations were made by Jefferson et al. (2004) and Eriksson et al. (2002) who highlighted the possibility to reduce smell associated with wastewater in applying chemical products. Another alternative is using underground greywater conservation tanks, which limit (neutralizes) the effects that the smell has on the household and the neighborhood. The positive coefficient of ASC in the base model is statistically significant, suggesting a significant status quo effect. The results show that ease of use is significant at 1% with expected sign of the coefficients.

In both the CL and MNL interactions model, Easyuse, Badsmell, and Disease are statistically insignificant. Lifetime is statistically significant and has the expected sign of the coefficients. In these models, the coefficients on choice attributes represent the preferences of base-case responders. By assumption, the base-case might represent responders who prefer a technology with a relatively longer lifespan, that is easy to use, that has a very low likelihood of generating bad smell and diseases at an affordable cost. As with the base-case models, the interactions models have a positive and significant ASC, indicating that responders had a preference for the status quo option, regardless of the change in the levels of the attributes. These results, therefore, show the importance of key attributes in explaining adoption of greywater technology within households.

The coefficients of the interaction terms describe the effects that individual characteristics have on preferences for each attribute. The significant negative coefficient on Innovators × Easyuse, Gender × Disease, Waterbill × Easyuse, and Income × Cost indicate that early adopters prefer technologies that are less difficult than late adopters. Women are more sensitive to technologies that have a high probability to generate diseases than men. Responders who believe that greywater helps to reduce the water bill prefer relatively easy technology than their counterparts. The significant and positive coefficient for Support × Cost, Gender × Cost, Waterbill × Badsmell, and Income × Lifetime indicate that greywater supporters are less sensitive to changes in costs associated with investment in water-saving technologies than non-supporters. Men are less sensitive to costs than women; those who believe that the use of greywater technologies actually reduce a water bill do not care much about bad smell than others. Those with higher incomes, however, seem to care about the lifespan of the technology than responders with lower incomes.

5.2 Latent class model

The specification of the LCA model was performed using all the variables that appear in the CL and MNL interactions models. Out of all model specifications, a three-class model was selected as the best specification based on AIC and BIC criteria. As shown in Table 7, the coefficient on Badsmell, Cost and ASC for class 1 are statistically significant. The coefficients for Easyuse and Cost are significant for class 2, while the coefficient of Easyuse, Badsmell, Disease, and Cost are highly significant for class 3. There is consistency on the sign of the coefficients for Easyuse, Badsmell, and Diseases for all classes except for Cost in class 3.

Class 1 has the largest share of responders with 89% and by default it is used as the reference class. This class is labeled as risk averse. Determinants of class membership for the other alternative classes are interpreted with respect to class 1. Class 1 characterizes responders who think that adoption of greywater technology would contribute to an increase in diseases and bad smell. These respondents are labelled as risk averse, because they are more likely to be reluctant to adopt greywater given the potential negative health-related consequences on their livelihoods. The class has the youngest (52 years) members and relatively more women (41%). Our study supports other findings that show women to be more risk averse than men (Charness and Gneezy 2012). We find that women have a strong preference for change away from the status quo (negative and significant ASC). The mean willingness to pay for Badsmell and Diseases are higher for class 1, suggesting that members of this class put more weight on badsmell and diseases when selecting water-saving equipment. Relative to classes 2 and 3, members of class 1 are significantly less likely to adopt technologies that lead to badsmell and diseases. Relative to members of class 2, they are also significantly less likely to innovate or to lead the technology adoption process.

Class 2 has the least class membership and accounts for 10% of responders. Members of class 2 are the innovators and leaders in using water-saving technologies. By definition, these are the members who are likely to adopt greywater technologies ahead of others. Members of class 2 are mature (56 years), slightly more educated (13.7 years of school) than the sample mean and composed of fewer women (26%) compared to other classes. The household marginal MWTP for Lifetime (527.79) and Easyuse (-1426.69) are higher for class 2, suggesting that members of this class care about the lifespan and ease of use of the technology (Table 8). The ASC is not significantly different from zero, suggesting the absence of status quo bias, thus, members in class 2 are indifferent between the status quo and other alternatives. This seems to suggest that members of class 2 might view greywater technologies as an appealing water-saving strategy if benefits are greater than the costs of adopting the technology.

Table 8 Household marginal willingness to pay per month, by class
Full size table

Class 3 is the second largest class, representing 85% of responders. They have been labeled “greywater supporters.” Class 3 characterizes responders who support initiatives that facilitate adoption of greywater technologies. The responders in this class also think that adoption of greywater technology would contribute to reduced water scarcity and postpone occurrence of Day Zero. For that class, adoption of water-saving equipment is naturally perceived as a tangible option that could help mitigate water crisis alongside other water-demand management options. Members of class 3 are mature (54 years), and women constitute 37% of the group. Again, the ASC is not statistically different from zero, suggesting the absence of status quo bias. This suggests that members of class 3 were indifferent between the status quo and moving away from it. Class 3 has the least MWTP for all attributes (Table 8). The positive and significant cost coefficient in this class suggests that responders were ignoring the cost associated with each alternative or selecting alternatives with higher cost with all else held constant. Although this is not consistent with expectations, the explanation could be that greywater supporters were willing to adopt any technology that can reduce household demand for freshwater to prevent a water crisis. This may be because the publicity provided by officials about the likelihood for the city to run out of water has contributed to rising concerns about future water crises.

5.3 Marginal willingness to pay and aggregate willingness to pay

When we consider the marginal willingness to pay (MWTP) across models, we observe that both the CL and MNL base-case models have the highest MWTP for Lifetime, Easyuse, and Badsmell, while the LCA has the highest MWTP for Disease. Mean MWTP for Lifetime, Easyuse, and Badsmell are higher in the CL/MNL base model, while the mean MWTP for Disease is higher in the MNL interactions model. Overall, Lifetime has the largest MWTP, followed by Easyuse, and Badsmell. Table 8 shows the willingness to pay across all the three classes. Respondents from class 1 have a positive willingness to pay for bad-smell and disease. This shows the high value they place on the negative externalities associated with usage of wastewater within urban areas. These are the respondents who exhibit high risk aversion again greywater technology given the potential health-related externalities that emanate from usage of recycled wastewater within premises. They are willing to pay an average of (1087 ZAR) and (735 ZAR) for bad-smell and diseases, respectively, to avoid these externalities. Respondents from class 2 have positive willingness to pay for lifetimes and easy use. They value reliability of the technology. This result is in line with the characteristics of the respondents from class 2. They have the higher likelihood of finding new innovative ways to neutralize the negative externality associated with usage of greywater. What is more important is adoption of wastewater conservation practices that contribute to reducing the expenditures on water bills, which in turn is expected to prevent water scarcity. Respondents from class 3 exhibit negative willingness to pay for all attributes but disease.

When comparing LCA model with CL and MNL models, we found that LCA provides a better fit to the data. Table 9 shows the willingness to pay by comparing LCA with CL. All the attributes carry negative signs in the LCA model. The CL base model has positive signs for all attributes but Disease, while the CL-interaction model has negative signs for Easyuse and Disease, and positive signs for Lifetime and Badsmell. All the models converge on the negative sign of Disease. The relatively low value placed on bad smell suggests that responders seem not to consider bad smell to be a very important attribute when making a choice about technology adoption. This is particularly true for greywater technologies that are adopted by households because of the fact that the technology helps reduce freshwater demand and hence the water bill. The technology has mitigation strategy during periods of water scarcity such as droughts.

Table 9 Average monthly household MWTP (LCA mean, conditional-base and conditional interaction)
Full size table

6 Conclusion and policy implications

This paper investigated the factors that drive adoption of greywater treatment technology, as a water-saving technology, using a novel econometric technique that accounts for household heterogeneity. The results show that households are sensitive to disease and duration of the technology on top of all the provided attributes, when it comes to explaining the attributes that determine adoption. Interestingly our results show that women are more sensitive to technologies that have a high probability to generate diseases than men. This means female heads of households in our surveyed areas are more cautious about the health status of their children and other family members than men. When running our LCA, almost all the attributes are significant. Based on the goodness of fit of the model, three classes were formed: risk-averse (class 1), innovators (class 2), and greywater supporters (class 3). Class 1 has the largest class membership and accounts for 89% of responders. In class 1, bad smell and ease of use are significant. Class 2 has the least share in our sample size (10%), reflecting the limited diffusion of effective technology within urban households in Cape Town. Ease of use and the cost of the technology are significant within that class. Whereas for class 3, ease of use, bad smell, disease, and costs, are all significant.

Our results show the importance of better understanding the profile of the potential users to promote widespread adoption of greywater technology. Policy interventions may support initiatives that attempt to design better greywater treatment technologies, which control for smell and eliminate any possibility of risk from waterborne diseases. This means that a massive investment in research and development should be promoted around greywater technology advancement. Alongside these technical interventions, our results show the importance of raising awareness, via public campaigns, about attributes that affect greywater technology. The benefits of such public campaigns arise from a better explanation of the real attributes of the technology in order for potential users to differentiate between the true advantages associated with decentralized wastewater treatment and the fake news provided in some media outlets. Finally, costs may also hinder adoption of greywater technology. Policy interventions may be articulated around possible financial support that could assist poor households in acquiring such technology.