2.1 Schistosomiasis as a public health problem: what are the potential benefits of interventions?

In endemic communities, the parasite S. mansoni is particularly difficult to avoid; individuals are at risk of infection simply by contact with infected water, for example whilst performing daily routines such as fetching water, washing clothes, bathing or whilst working (e.g., fishing). In these poor rural communities, there is a severe lack of sanitation (WHO, 2013), meaning that even if an individual knew how to reduce disease transmission, they are not always in a position to do so. We see two distinct categories where WASH interventions are capable of being effective. First, improved safe water access can limit community dependence on infected bodies of water, leading to reduced exposure time and thus reduced risks of (re)infection to individuals. Second, improved sanitation may reduce the number of eggs that reach fresh water to continue their life cycle. In a review of studies from Africa, Asia and Brazil (the main affected country in South America), Grimes et al. (Reference Grimes, Croll, Harrison, Utzinger, Freeman and Templeton2014) found that access to improved water and sanitation significantly lowered the likelihood of schistosome (re)infection.

We thus classified potential WASH interventions at two discrete points in the parasite life cycle: Risk to Self or RTS – ‘improved water access’; and Risk to Others or RTO – ‘improved sanitation and use’. Exposing oneself to infected water (e.g., collecting water for washing or cooking) is categorised as a behaviour that puts oneself at higher risk. Open defecation puts the whole community, including oneself and others, at risk, as the faeces will then have a higher chance of reaching freshwater, sustaining the parasite lifecycle. In our experimental design we accounted for these two distinct behaviour groups related to two different aspects of risk-reduction benefits: mainly private risk reductions in the case of RTS versus mainly public risk reductions for RTO. We anticipated that: (i) WTP and WTW measures would vary according to whether a RTS or a RTO intervention was in question; (ii) that the effect of higher money or labour prices on choices would vary between RTS and RTO; and that (iii) different factors would explain the variation in WTP and/or WTW for mainly-private benefits as opposed to mainly-public benefits.

2.2 Previous literature

Many stated preference studies related to health in LMICs are centred around water quality and accessibility or improved health care services (e.g., Whittington et al., Reference Whittington, Davis and McClelland1998; Nam and Son, Reference Nam and Son2005; Baltussen et al., Reference Baltussen, ten Asbroek, Koolman, Shrestha, Bhattarai and Niessen2007; Youngkong et al., Reference Youngkong, Baltussen, Tantivess, Koolman and Teerawattananon2010; Abramson et al., Reference Abramson, Becker, Garb and Lazarovitch2011; Vásquez, Reference Vásquez2014; Gibson et al., Reference Gibson, Rigby, Polya and Russell2016). Limited economic valuation research has been performed that elicits preferences for more extensive WASH interventions. Our study adds to the DCE literature in LMICs by asking respondents not just about water quality, but also about other health practices, specifically preferences for improved WASH.

There are particular challenges of implementing a DCE in low-income communities. Mangham et al. (Reference Mangham, Hanson and McPake2009) and Bennett and Birol (Reference Bennett and Birol2010) outline some of these, including assigning attributes, including non-monetary payment vehicles (e.g., Abramson et al., Reference Abramson, Becker, Garb and Lazarovitch2011; Rai and Scarborough, Reference Rai and Scarborough2012, Reference Rai and Scarborough2015; de Rezende et al., Reference de Rezende, Kahn, Passareli and Vásquez2015; Gibson et al., Reference Gibson, Rigby, Polya and Russell2016; Hagedoorn et al., Reference Hagedoorn, Koetse, Van Beukering and Brander2020), finding appropriate survey administration modes, and low literacy levels amongst respondents (see also Christie et al. (Reference Christie, Fazey, Cooper, Hyde and Kenter2012)). To address these issues, we used attributes identified by stakeholders from the communities after several months of ethnographic research in each of the three communities, as described in more detail elsewhere (Ssali et al., Reference Ssali, Pickering, Nalwadda, Mujumbusi, Seeley and Lamberton2021), administered the questionnaire in person through local enumerators fluent in the local language, and used two payment vehicles.

We included both labour and money as payment vehicles in each of the choice cards, to address payment vehicle issues. Furthermore, in designing our DCE we incorporated a bottom-up approach. Bottom-up approaches embed local stakeholder strategies directly into the decision making process (Carolus et al., Reference Carolus, Hanley, Olsen and Pedersen2018). All attributes and levels evolved directly from stakeholder suggestions, which further helps to create a sense of ownership and enhances future successful policy development.

Our study adds to the previous work of Meginnis et al. (Reference Meginnis, Hanley, Mujumbusi and Lamberton2020) which examined data from the RTS choice experiment only. Their aim was to explore preferences for RTS intervention, but more specifically to examine the impact of using both monetary and non-monetary numeraires in DCEs. They calculated the shadow wage rate between labour and money to be between 15–55 per cent of the market wage rate. However, the models presented in Meginnis et al. (Reference Meginnis, Hanley, Mujumbusi and Lamberton2020): (i) do not contrast interventions which target RTS with those that target RTO; and (ii) fail to address the large WTP and WTW values their models yield, which far exceeded the highest levels of monetary and labour contributions of the choice cards. As such, our paper's contribution is both a more robust set of policy-relevant valuation estimates through an Equality Constrained Latent Class Analysis (details in section 4), and a contrasting of community preferences for RTS as opposed to RTO interventions with direct public-health policy relevance.

3.1 Risk to others design

The RTO choice cards consisted of interventions associated with providing new public latrines, cleaner latrines and introducing fines to deter open defecation. We included five attributes: (1) location of new public latrines; (2) latrine maintenance; (3) open defecation fines; (4) monthly monetary cost; and (5) weekly labour contributions.

The location of new public latrines was described to be either at the lake shore, the market place, or in a residential area located within 5 min' walk from one's household. We also included a fourth level for ‘no new public latrines built’. The second attribute was for the maintenance of latrines. From the three months of REA we undertook to inform the DCE design, it was discovered that one of the largest deterrents to wider use of the few public latrines already found in the villages is their poor condition. As such, many members of the community described the public latrines as too dirty for them to want to use. In order to persuade the community to use new public latrines, we expected that a certain cleanliness standard must be maintained. We therefore included an attribute describing whether public toilets (new and existing) would be maintained at the current standard (labelled ‘current’) or that all public toilets would be maintained regularly to a high standard of cleanliness (labelled ‘improved’).Footnote ²

In order to deter open defecation in line with current and past strategies identified during the ethnographic appraisal, we included an attribute for fines issued to individuals caught openly defecating. We suggested a fine be in place for open defecation within 30 m of the lake.Footnote ³ Alternatively, a fine for open defecation anywhere in the community could be employed, or no fine. Finally, we included two payment vehicles: monthly payment and weekly labour. Both payment vehicles were described as being payments per household and would go towards construction, maintenance and organisation of the new investments. Table 1 summarises the attributes and levels for the RTO choice cards.

Table 1.

Attributes and levels for RTO discrete choice experiment

3.2 Risk to self design

Individuals are put at risk of contracting schistosomiasis when they wade into Lake Victoria to collect water for domestic chores (e.g., cooking, cleaning, laundry) and personal hygiene (e.g., bathing). Water is also used for washing young children in basins, cleaning motor bikes, fishing, farming, and making bricks, activities which further endanger individuals. The RTS choice cards have five attributes related to water use: (1) water source; (2) water access; (3) education campaign; (4) monthly monetary cost; and (5) weekly labour contributions.

In order to reduce the reliance on potentially infective freshwater, we included a ‘new water source’ attribute, with four improved levels. As safe tap water is currently paid for in the communities, people tend to limit their use for drinking as it is too costly to substitute this water for all water needs.Footnote ⁴ We therefore included two levels with a new tap to be built where households can either fetch two 20 litre jerry cans a day, or where households can collect ten 20 litre jerry cans a day. We hypothesised that respondents would prefer a tap with a lower restriction if they envision using tap water for all water needs, not just for drinking. Two additional levels for a new water source were also included. One was described as a sun filtration centre where water is fetched from the lake and then brought to the centre and exchanged for water which is safe for domestic chores but not for drinking. Evidence suggests that leaving infected water in the sun for 24–48 h kills any schistosome cercariae and is therefore safe to use with respect to transmission of catching bilharzia, although it is not safe to drink (Braun et al., Reference Braun, Grimes and Templeton2018; Morse et al., Reference Morse, Luwe, Lungu, Chiwaula, Mulwafu, Buck, Harlow, Fagan and Mcguigan2020). We specify an additional attribute level with a filtration centre that exchanges lake water for water which is also safe for drinking, using a more sophisticated filtration mechanism.Footnote ⁵

The second attribute concerned improved access to natural bodies of water. Individuals currently wade into the lake to around mid-thigh in order to reach water deep enough to submerge their jerry cans to fill them up. Water is collected for consumption (e.g., drinking and cooking) but also for bathing, household chores (e.g., laundry), washing motor bikes, farming, etc. Time spent in the lake, as well as body surface area exposed to water, both increase the risk of infection. Furthermore, wading into the lakes (e.g., to enter boats) also exposes people to infectious water. We therefore proposed new water access points, such as piers, that extend into the lake above the water. Individuals could walk along these and access lake water by leaning over and filling up their jerry can, reducing the contact surface area as well as the total time in the water.

A remerging theme in the initial ethnographic appraisal focus groups and interviews was the desire for more health education campaigns. There is mixed evidence on whether education campaigns have any effect on behaviour change, but there is some indication that education at least affects awareness (Price et al., Reference Price, Verma and Welfare2015). We included three new levels concerning an education campaign, each varying the way information is presented. In these villages, past campaigns have painted murals in prominent places that use images to encourage healthy behaviour. We included a level for a mural education campaign and a level for a public radio campaign. This would utilise pre-existing loud speaker systems that play daily and would communicate safe WASH behaviours. A third level provided monthly community talks led by village health teams; individuals would need to congregate at an agreed space and listen to the presentation, participate in an activity, etc.Footnote ⁶

Finally, we included the same two payment attributes as the RTO. Table 2 outlines the attributes and levels for the RTS choice cards.

Table 2.

Attributes and levels for RTS

A D-efficient design was generated for both the RTO and RTS choice cards in Ngene (ChoiceMetrics, 2018). Restrictions were imposed such that attribute levels that in combination did not make sense were not shown on the same alternative (for example, if contributing a monthly fee, alternatives had to have at least one improved attribute). Similarly, all new investments were restricted to either cost in money, time or both: there could not be an intervention that was completely free. Each choice card had two policy interventions with a positive labour and/or money cost, and a third option where no new interventions were implemented at zero additional money/labour cost. As most members of the communities surveyed had only limited literacy, pictures rather than text were used in the choice cards. In each choice card, enumerators described every alternative by explaining the provisions shown through the pictures. The third option was always presented as creating no new interventions. An example choice card for the RTO and RTS choice sets can be seen in figures A1 and A2 of the online appendix.

3.3 Data collection

A total of 15 choice tasks were split into three blocks for each of the RTS and RTO experimental designs.Footnote ⁷ Respondents were asked to complete five RTS choice tasks and five RTO choice tasks.Footnote ⁸ All respondents therefore answered ten choice tasks, which is at the upper limit of what is recommended to not lead to increased respondent variance (Bech et al., Reference Bech, Kjaer and Lauridsen2011) and kept the survey interview time under one hour in length.

The order of RTS and RTO tasks was randomised such that half the respondents saw the RTS first and the other half saw RTO first. After the DCE tasks, respondents answered questions about their behaviours concerning water use and sanitation practices. In line with recent literature on consequentiality in stated preference studies (Zawojska et al., Reference Zawojska, Bartczak and Czajkowski2019), we asked respondents to answer Likert-scale questions regarding the likelihood that: (i) the survey will actually be used to affect policy interventions; (ii) payments will actually have to be made; and (iii) labour will actually have to be contributed. We also asked demographic and socio-economic questions concerning daily income, household size and education.

We surveyed 425 individuals in total. Respondents were surveyed in each village proportional to their village populations, with the largest sample coming from Bwondha (n = 247), then Bugoto (n = 130; n = 85 from Bugoto A and n = 45 from Bugoto B), followed by Musubi (n = 45). Respondents in each village were randomly surveyed according to a sample interval based on the desired sample size per village and the estimated number of households in that village. Sample characteristics are summarised in table A1 of the online appendix.

4.1 Equality constrained latent class model

The MNL model assumes that every respondent considers every attribute (as well as assuming that preferences do not vary across the sample for each attribute). However, it is increasingly thought that respondents in choice experiments make use of simplifying decision heuristics, given the cognitive burden of fully rational choices. One such heuristic is to ignore one or more attributes when making choices, a phenomenon referred to as attribute non-attendance (ANA) (Campbell et al., Reference Campbell, Hensher and Scarpa2011; Koetse, Reference Koetse2017). Given the possibility of this happening with our respondents, we considered an additional Equality Constrained Latent Class (ECLC) model for both the RTO and RTS scenarios.Footnote ¹¹ In order to compare the two risk domains, our analysis hinges on comparing MWTP and MWTW values; it is therefore imperative that our estimations are not inflated by artificially small (i.e., zero, or close to zero) estimations of ${\delta _{pi}}$, which would inflate W_k in equation (5).Footnote ¹² We therefore specifically explored payment (both money and time) ANA.

To estimate an ECLC model we assume that there are four types of respondents: people who fully attend to all attributes, and three possible ANA cases. These restrictions are such that attributes which we hypothesised respondents ignored are restricted to zero. The parameters for the attended attributes are held constant across classes, such that we are accounting only for attribute non-attendance, not preference heterogeneity between these latent classes of respondents (Koetse, Reference Koetse2017; Glenk et al., Reference Glenk, Meyerhoff, Akaichi and Martin-Ortega2019).

We explore only payment-ANA, and do not consider other forms of attribute non-attendance. This is because it is difficult to separate ANA from attribute non-importance. However, attribute non-importance is unlikely to be the case of the payment vehicle as individuals are expected to always care about the cost of an item, despite not attending to it in a specific choice context. Therefore, for the payment vehicle, non-attendance and non-importance are theoretically unconfounded (Koetse, Reference Koetse2017). We focused on three possible ANA strategies: ignoring the monthly monetary costs, ignoring weekly labour requirements, or ignoring both. As such, we have four possible classes, and the probability that respondent n chooses alternative i is now conditional on being in class s and can be written as:

(6)\begin{equation}{P_n}({i|s} )= \frac{{exp ({\beta^{\prime}{x_{ik}} + {{\delta^{\prime}}_{is}}{c_{is}}} )}}{{\mathop \sum \nolimits_j exp ({\beta^{\prime}{x_{jk}} + {{\delta^{\prime}}_{js}}{c_{js}}} )}}.\end{equation}

The utility of the K non-cost attributes and the ASC ($\beta$) are kept fixed across the classes.Footnote ¹³ For classes 2–4, the respective payment attributes are constrained to exhibit a zero fixed marginal utility. The probability that an individual n is in class s is expressed through a standard multinomial logit model:

(7)\begin{equation}{\pi _s} = \frac{{exp ({{\theta_s}{z_n}} )}}{{\mathop \sum \nolimits_{s = 1}^S exp ({{\theta_s}{z_n}} )}},\end{equation}

where ${z_n}$ represents individual co-variates to include in the class membership model. We consider several individual covariates in our class membership model, outlined in the online appendix, section 7. Marginal WTP and WTW for the ECLC models are estimated exactly as before in equation (5), except that they should be weighted according to the percentage of respondents that attended to the payment attributes.