We followed the Bridges et al. checklist for conjoint analysis in health,18 and report against it in Supplementary Material A. Data and replication code are available at https://osf.io/38vsh/
Study population and sampling
The study took place in two cities in Mozambique: Maputo (population 1.1 million) and Dondo (population 100,000). Since this is the first time DCE-based valuation has been used for any index of sanitation-related quality of life, our sampling priority was not to achieve representativeness of the cities’ population. Rather, we aimed to achieve approximate gender balance and diversity in type of toilet used (Supplementary Material B). The study population was adults aged 18 + living in two neighbourhoods/bairros in Maputo (Polana Caniço A and Polana Caniço B) and two in Dondo (Macharote and Nhamainga). These areas were selected because they are mixed in terms of housing quality and, in particular, have a diversity of toilet types used.
The majority of healthcare DCEs interview 100–300 respondents.19 We aimed to recruit 600 respondents to meet other study objectives, as well as to allow for dropping some of the sample on data quality grounds.20 We aimed to interview 300 women and 300 men per site, to allow exploration of whether valuation varies by gender. For toilet type, we aimed to sample 200 people using flush toilets, 340 people using pit latrines, and 60 people practising open defecation (no toilet). We achieved this by sampling based on data from existing health surveillance surveys (details in Supplementary Material B).21 The bairros in Dondo were selected on the basis of surveillance data indicating that prevalence of open defecation was > 10%, since open defecation was uncommon in the Maputo site (Supplementary Material B).
SanQoL-5 index
The SanQoL-5 is a multi-attribute measure of sanitation-related quality of life, developed from primary qualitative research and supported by the literature on what people value about sanitation.3,12 Its descriptive system (Table 1) has five questions, each measuring a capability-based attribute: disgust, disease, privacy, shame and safety. Each is measured on a three-level frequency scale (always, sometimes, never), with questions framed such that “never” is the best outcome. There are therefore 15 attribute levels to be valued in 243 (= 35) possible combinations. Following norms in health-related quality of life (HRQoL), each combination is termed a “sanitation state”. Adopting HRQoL notation, the best state is denoted 11111 (“never” for all levels) and the worst 33333, with intermediary states such as 23132, 11213, etc.
Table 1
SanQoL-5 descriptive system
Attribute | Question* | Responses |
Disgust | How often do you feel disgusted when using the toilet? | Always Sometimes Never |
Disease | How often do you worry that the toilet spreads diseases? |
Privacy | How often do you worry about being seen while using the toilet? |
Shame | How often do you feel ashamed about using the toilet? |
Safety | How often do you feel unsafe while using the toilet? |
* A preamble is as follows: “The following questions are about your sanitation experiences in the past 30 days, meaning defecation, urination, and anything else you do in a toilet. Please respond with always, sometimes or never.” If less literate respondents struggle with a question, it can be reformulated as “Do you feel disgusted while using the toilet? How often?”. Before the SanQoL-5 questions, the respondent is asked about the last place they defecated. If the respondent practiced open defecation (OD), e.g. in fields or wasteland, they are directed to OD-specific questions, e.g. “How often do you worry about being seen while practising open defecation?” |
Data collection
Our study was a face-to-face survey using Open Data Kit (ODK) Collect software on tablet computers. Though the survey was administered in Portuguese in the vast majority of cases, some participants preferred to speak in the predominant local language (Changana in Maputo, Sena in Dondo). Therefore, two teams (one per site) were recruited and underwent five-day programmes of training and piloting. Data collection was undertaken during May-July 2023. The questionnaire was translated into Portuguese by NB and the translations discussed at length with field team. No incentives for participation were provided.
Discrete choice study design
After questions about socio-economic status and sanitation, the DCE section started with a series of warm-up tasks, to ensure participants fully understood the choices they were being asked to make. First, participants answered the SanQoL-5 questions (Table 1) and completed the sanitation visual analogue scale (VAS) – a 0-100 scale on which people rate how they feel about their level of sanitation today (Supplementary Material B). Second, participants watched three video vignettes on the tablet, to provide more meaning to hypothetical states and introduce the images used to frame attributes (Supplementary Material B). In each video, a hypothetical person describes the toilet they use and how it makes them feel about each of the SanQoL-5 attributes, i.e. describes their sanitation state. After each video, the participant was asked to score that person’s level of sanitation on the VAS, to get them used to the idea of comparing states. Third, participants were asked to complete a food-based menu choice card (Supplementary Material B), to emphasise that the two columns as a whole are being compared and trading items between columns was not possible. The last warm-up task involved being shown three sanitation states, and asked to choose which was worst and which was best, as well as their reasoning, to assess whether they understood the task (Supplementary Material B).
For the actual DCE choice tasks, participants were shown a card with two sanitation states as profiles of SanQoL-5 atttribute levels (Fig. 1) with the same emoji visualisation as the warm-up tasks. Participants were asked to select which state was “better”, with no opt-out. This follows best practice from valuation protocols for the EuroQoL 5-dimension (EQ-5D) measure of HRQoL.17,22 Participants were told to not consider their present toilet or level of sanitation, but instead to imagine being in the states in the scenario.
Each participant undertook 10 choice tasks. We identified a 6-block efficient design using the dcreate programme in Stata 18 with a Modified Federov Algorithm (d-efficiency 10.4).23 With 60 choice tasks (6x10), there were 120 sanitation states compared in all. Each block including states across the range of severity. To avoid bias from the ordering of the tasks (e.g. less care taken over later tasks) we randomised participants into 12 groups, with half of the groups doing tasks in reverse.24
Quality Control
Each interviewer undertook 8 pilot interviews (80 DCE tasks per interviewer in total), in areas outside the study sample – the data were not included in the analysis. The pilot identified issues with time taken to complete the survey, which resulted in removing some socio-demographic questions and randomising some non-DCE questionnaire modules to sub-samples. Overall, the acceptability of the tasks was good. Preliminary analysis of the pilot data indicated that DCE data were consistent. Other aspects of quality control included timestamps throughout the survey, to allow flagging when a participant completed a section extremely rapidly relative to most others. We also included a dominance test halfway through the DCE, in which one state (12121) was objectively better than the other (23232) on all five attributes, so there is a “correct” answer. Dominance test choices were not included in the analysis.
For the primary analysis, we excluded data of participants who met one or more of these conditions: (i) failed the dominance test; (ii) completed the first five tasks in less than 10 seconds per task; (iii) completed the second five tasks and dominance task in less than 5 seconds per task. We also examined the choices in respect of level sum score (LSS), which is the sum of attribute levels in state notation (e.g. the LSS of 11113 is 7). We calculated difference in LSS between the two options a respondent was shown and observed the distribution of responses. Hypothetically, a larger LSS difference should increase the likelihood that a respondent chooses the option with the lower LSS. The LSS of the best state (11111) is 5 and the worst state 15. We reconfirmed fieldworkers’ classifications of toilet types by verifying photos they took of toilets’ interiors against entered data on toilet characteristics.
Data Analysis
We analysed choices in Stata 18 first using a conditional logit model, which assumes that preferences are not correlated across individuals. We then using a mixed logit model with correlated parameters, which aims to account for: (i) preference heterogeneity (when differences between individuals’ preferences cannot be explained by observable characteristics); and, (ii) scale heterogeneity (when unmeasured factors affect individuals’ responses to different extents).25 We based model selection on whether there was evidence of heterogeneity, as well as the Akaike/Bayesian information criteria (AIC/BIC).
Our analytical approach was based on the EQ-5D valuation protocol.17,22 The model assumes that in making their choice, people are comparing the quality of life they would have in two sanitation states (Eq. 1), namely \(\:{V}_{ijl}\) (left-hand option \(\:l\) for individual \(\:i\) within DCE pair \(\:j\)) and \(\:{V}_{ijr}\) (right-hand option \(\:r\)). Eq. 1 represents the choice as an inequality, with the sign (< or >) decided by the participant’s response. Since there is no opt-out, the respondent cannot give them equal value.
Equation 1
$$\:{V}_{ijl}=\:\alpha\:-\:{\sum\:}_{k=1}^{10}{\beta\:}_{k}{x}_{k}^{ijl}+\:{e}_{i}^{lj}\:\:\:>\:?<\:\:\:\:\:{V}_{ijr}=\:\alpha\:-\:{\sum\:}_{k=1}^{10}{\beta\:}_{k}{x}_{k}^{ijr}+\:{e}_{i}^{rj}$$
The variable \(\:x\) represents a sanitation state using 10 dummy (binary) variables. The first two dummies refer to “sometimes” and “always” levels of the disgust dimension. If the state involves being “sometimes” disgusted (Table 1), the “sometimes” dummy takes the value 1. For the “always” level its respective dummy takes the value 1. If both dummies are 0, then the state include “never” being disgusted. The other 8 dummies are the equivalents for the remaining attributes.
If all 10 dummies are zero then the state is 11111 (full sanitation capability), its value denoted by α, which cancels out once the participant makes their choice. The parameter β is a 10 × 1 vector aligning to the dummies. Its first two elements reflect decrements of “sometimes” or “always” being disgusted against the value of “never” being disgusted. Since estimated coefficients are decrements, they are expected to be negative. The overall decrement of moving from “never” to “always” is the sum of the coefficients for “sometimes” and “always”. Error terms are assumed to follow an extreme value distribution. We rescaled estimated coefficients to a 0–1 index, whereby 0 is the value of the worst state and 1 the value of the best. This is achieved by dividing through by the sum of the coefficients of the worst levels (i.e. always).
We included a sub-group analysis by sex. First, we explored whether differences in preferences between women and men were explained solely by differences in randomness of choices by sub-groups, i.e. scale heterogeneity.26 We assessed this using the Swait-Louviere test,27 effectively a likelihood ratio test comparing the log likelihood statistics of the pooled model with sub-group models. Following Mott et al.25, our main sub-group analysis was to compare relative attribute importance (RAI) in the two sub-group regressions. To estimate RAI scores, we calculated ratios of each attribute’s “always” coefficient to that of the lowest-valued attribute. Attributes with higher RAI therefore have a higher value. We estimated RAI difference by subtracting RAI scores in the men’s sample from those in the women’s sample, and estimated confidence intervals using nlcom in Stata 18.
Ethics
The study received prior approval from the Comité Institucional de Ética at the Instituto Nacional de Saúde in Mozambique (ref: 028 /CIE-INS/2023), and the Research Ethics Committee at the London School of Hygiene and Tropical Medicine (Ref: 28190). Informed, written consent was obtained from all participants.