Protocol for a discrete choice experiment: understanding preferences for seeking health services for survivors of sexual violence in Uganda

Aims

The proposed study will address the above-described needs through the following objectives:

• Aim 1: Conduct formative qualitative research to understand factors that influence care-seeking behaviours of sexual violence survivors who have and have not sought formal care in order to inform attribute selection for a DCE.

• Aim 2: Using a DCE approach, systematically assess the factors that influence women’s decisions to seek services after experiencing sexual violence.

• Aim 3: Based on findings from the DCE, apply user-centred design principles to co-design an intervention that addresses barriers and preferences.

Findings from this study are intended to inform efforts to increase rates of reporting of sexual violence and improve reproductive health, mental health and related health outcomes for women in Uganda and, more broadly, in low-resource countries where rates of sexual violence are high.

The methodological details of our DCE study (Aims 1 and 2) are described in this paper.

Qualitative research to inform the DCE

We will first undertake preliminary qualitative research to identify the critical concerns of key stakeholders and define the scope of attributes and levels to be incorporated into the final DCE design. Identifying DCE attributes requires a deep understanding of the target population’s perspective and experience.55–57 Attributes must be salient, well-defined and understandable to respondents, and the levels must be relevant, capturing a range of meaningful outcomes and including both less and more desirable variations. Qualitative interviews have been identified as an effective method to clarify attributes and attribute levels for DCEs and, in undertaking this work, we will improve the face validity of our DCE.53 58

Our study capitalises on our ongoing partnership with the local village executive committees—and especially the secretary for Women and Children Affairs. The village executive committee assists with maintaining law and order; initiates, encourages, supports and participates in self-help projects; and mobilises people, materials and technical assistance for self-help projects. The village courts have the jurisdiction to handle cases relating to assaults and battery, conversion, damage to property, trespass, small debts, disputes with respect to customary land and civil disputes governed by customary law (customary land, marriage, divorce and inheritance).59

The village council secretary for Women and Children Affairs has the power to mediate and resolve issues related to the safety and welfare of children and typically serves as a confidant and arbitrator in marital affairs. Village secretaries are elected for political appointments based on community perceptions of these women as being trusted, approachable and natural helpers with the ability to assist women, children and families. By engaging the secretary for Women and Children Affairs, we will gain access to an otherwise hidden and hard-to-reach population of female survivors who may or may not interface with the health service systems. Our approach will also reduce potential bias with the recruitment of women at clinics and other locations, such as police stations, by engaging women who may not interface with formal sexual violence services. Village secretaries are often consulted by women and/or families who reach out to them directly for advice or to aid in private conflict resolution. As such, they are more attuned to cases of sexual violence in their community and are trusted not to perpetuate community gossip or other forms of information sharing. Our approach is consistent with numerous studies that engage natural helpers within communities to mobilise, recruit and engage hard-to-reach populations.

At the beginning of the study, we will convene a meeting for all secretaries where we will brief them on the study, including the study goals, eligibility criteria and our procedures for protecting human subjects involved in research. The secretaries will also receive modified training on good clinical practices that will emphasise the importance of respecting participants’ confidentiality, autonomy and privacy in research. The village secretaries will first contact potential participants to brief them about the study and assess their interest in participating. Subsequently, they will present potential participants who express an interest in participating in the study with two choices: (1) give verbal permission to the village secretary to share their contact information with the research team or (2) the village secretary will provide them with the research team’s contact information so they can contact the team directly. Following this contact, the research team will then reach out to each woman to verify their eligibility and then schedule a place and time to conduct the interview. We will ask these village secretaries to start recruiting women who have most recently experienced sexual violence. From this pool, we will recruit 56 survivors meeting the inclusion criteria: survivors who have experienced sexual violence in the last 12 months and used health services (n=28) and survivors who did not use health services (n=28). We will continue screening referred women until we realise the desired sample size. We will also interview each of the village secretaries. This sample size will be sufficient to reach theoretical saturation.60–62

Participants will be interviewed using semistructured interview guides (see online supplemental material 1 and 2), intended to elicit rich data on social norms surrounding sexual violence and perceptions and decision-making processes surrounding health-service seeking after experiencing sexual violence, including multilevel barriers and facilitators to seeking health services. The interviews will also include questions to identify and refine attributes for the DCE questionnaire and to elicit contextually appropriate attribute levels and descriptors relevant to survivors’ preferences for health services. Survivors will be asked to identify factors that influence whether, where and how they seek services, as well as the positive and negative features of service options.55 61 63 64 Interviews with village secretaries for Women and Children Affairs will also focus on eliciting barriers to and facilitators for health service utilisation at the individual and community levels.

Interview guides will be translated from English to Luganda and back-translated for accuracy. Interview guides will then be reviewed to ensure that the questions sound conversational and natural and will be revised accordingly. Finally, interview guides will be pilot tested to ensure the questions are culturally relevant and suited to the local vocabulary. Semistructured in-depth interviews will be conducted in a private location to be selected by the participant and will take approximately 90 min.

Interviews will be digitally recorded, transcribed and translated. All transcripts will then be uploaded to Dedoose for data analysis.65 All transcripts will be reviewed by the research team to develop a broad understanding of the content and identify topics of discussion and observation. Using analytic induction techniques,66 transcripts will be read multiple times for this initial coding by the research team. For initial coding, a random selection of 10 interview transcripts will be read multiple times and independently coded by the team using a priori (ie, from the literature) or emergent themes (also known as open coding).67 Broader themes and categories will be broken down into smaller, more specific units until no further subcategory is necessary. Analytic memos will be written to further develop categories, themes and subthemes and to integrate the ideas that emerge from the data.67 68 The codes and the inclusion and exclusion criteria for assigning a specific code will be discussed as a team to create the final list of codes (ie, a codebook). Each transcript will then be coded independently. If needed (based on the number of attributes identified), a stakeholder workshop will then be organised.

Refinement of attributes and levels for DCE and construction of choice sets

While the DCE will focus primarily on actionable features of the health system, qualitative interviews may reveal other important factors (eg, stigma, personal relationships and trust in the legal system). For example, confidentiality and personal relationships were key factors influencing HIV screening preferences in a DCE study in Tanzania. Social norms, cultural factors and networks are rarely employed as attributes in health DCEs.69 Such information will be probed and included as control variables and interactors in the DCE survey and as predictors of preference heterogeneity, such as latent class membership, when we examine variation in preferences. As appropriate, we will also test the inclusion of such factors as attributes directly in the DCE (see online supplemental material 3).

The number of attributes included in a DCE depends on the research question and application, the elicitation technique used and researcher preferences. In practice, most stated-choice DCEs contain fewer than eight attributes so that respondents can consider all attributes simultaneously when completing the survey. Our goal is a standard ‘full profile’ stated-choice DCE, so we will aim for a maximum of 6–8.70 We aim to ask respondents to consider up to nine pairwise choice sets, although this can vary based on the complexity of the choice sets and target population characteristics and will be assessed during piloting and survey refinement. We will also ask respondents a second question, known as an ‘opt out’,53 that allows them to indicate that, although they selected one of the two alternatives, they would select that health service if available, or if they would prefer neither option. This will allow us to capture data to estimate the unconditional demand for seeking health services after experiencing sexual violence. Including this option will more closely reflect the real-world context and will allow us to ascertain how other respondent characteristics will determine the unconditional demand for health services.

Once the DCE instrument has been finalised, the questionnaire will be piloted among 10 participants to identify any conceptual overlap between attributes, respondents’ understanding of the survey and the optimal number of choice sets to be included.

After finalising the attributes, levels, choice elicitation format and other aspects of survey design, the final step before fielding is to generate the experimental design or the specific questions and comparisons that are shown to respondents out of the numerous possible combinations of attributes and levels. Key objectives for the design include orthogonality (statistical independence),71 level balance (the range of options shown is roughly equal) and minimal overlap between alternatives. NGene V.1.372 will be used to generate experimental designs, following standard guidelines.72 73 The final design will be optimised for the intended statistical analysis (latent class or random parameters logit) based on the D-efficiency score.74–76 During fielding, the internal consistency of responses will be assessed by including one or two choice pairs in which one option is superior to the other on all characteristics (a dominated alternative).

Participants and sampling

The target population of this study centres on female survivors of sexual violence recruited through our network of village secretaries for Women and Children Affairs. We will verify eligibility for the study using the standardised gender-based violence screening questionnaire.77 Potential participants will be excluded from the study if they cannot comprehend the study and participant rights or refuse to participate. As part of the research process, all participants will be provided with a resource list of sexual violence support services within their community, at the regional level and nationwide. These resources will include legal services. If participants are screened by the research team and deemed ineligible, they will be thanked for their time, provided the same incentive provided to those invited for a full interview and their time with the study will be politely concluded.

Power calculations for DCEs are challenging, as they are a function of the efficiency of the experimental design, the statistical method used and the salience of the attributes and levels to the sampled population.78 For this study, we computed the sample size as: n≥500×(l÷(J×S)), where l is the largest number of levels for an attribute (including interaction terms), J is the number of alternatives in each choice task, and S is the number of tasks. Using l=10, J=2, S=8, typical parameters for this type of DCE, the minimum required sample size for the DCE is 312.78 79 Using previous estimates of various forms of sexual violence from Uganda, we will assume approximately 29% of women aged 18–49 years have experienced sexual violence in the past 12 months.28 As such, in order to ensure a sample size of at least 312 women for the DCE, we aim to recruit 20 women from each of the 39 villages (n=780) and randomly select 351 women from this participant pool. Allowing for a 10% non-response rate, the expected realised sample size will be 316 women.

Experimental design

Using the finalised DCE instrument, will assess not only the overall demand for health services among women who experience sexual violence in Uganda but also the factors that would influence their decision to seek services. Quantitative data will be gathered through a structured survey delivered using face-to-face interviewing and CAPI technology. Data will also permit estimation of women’s relative preference for specific health service-related attributes when choosing between different service packages. We will apply the DCE to assess the demand for health services designed to present respondents with a series of choices between hypothetical scenarios (eg, the choice between health service packages) comprised of preselected attributes (eg, cost, service quality and availability) at varying levels (eg, high vs low quality or varying costs incurred).

The DCE section of the full survey questionnaire will begin with a description of the attributes, levels and a simple introduction to choice-set questions. To minimise bias caused by the order in which the choice sets occur or the attributes are described, we will use several levels of randomisation (blocked designs or ‘versions’ of the survey, order of questions within a block, order of attributes within the survey and right/level ordering of specific tasks).80 Pictures, diagrams and symbols will be added when feasible to aid comprehension, which will be particularly relevant in this setting of limited literacy.57 The questionnaire will be translated into Luganda and administered using tablets by trained fieldworkers. The intended start date of the DCE is May 2024, with completion scheduled for November 2024.

Analytical plan

Sociodemographic and prior healthcare utilisation characteristics will be described using frequency tables. Descriptive analyses will also include univariate measures of central tendency and variability (eg, means, medians, ranges and SD) for continuous measures. DCE data analysis uses an analytic model based on random utility theory, which posits that an individual n’s true latent but unobservable utility U_i
for alternative I in a choice situation can be depicted as U_i
=V_in + ε_in
, where V_in
=V(X_in
, Z_n
) is the systematic component of the utility of individual n with individual characteristics (Z_n
) for a scenario with attributes (X_in
). ε_in
is an unobservable random component. Allowing βx_ni=V_in
, the probability of choosing I from J alternatives can be written as the standard logit formula:

Our primary model will be a mixed (random parameters) logic model, which has several advantages over simple conditional logit models. First, they allow preference (estimated coefficients) to vary across respondents, accounting for (often substantial) preference heterogeneity.80 Second, these models adjust estimates for repeated choices by the same individual on the survey and typically have a better model fit.79 For Aim 2, we will also model the role of individual-specific factors (eg, demographic characteristics, stigma and cultural norms) through the inclusion of interaction terms, individual characteristics and DCE attributes.52 We will also use a latent class logit model, specifically a discrete analogue of mixed logit, which examines statistical ‘group’ preferences by observing variation in stated responses. An advantage of latent class analysis is that we can predict class membership using individual characteristics to identify potentially different groups of participants, analogous to segmentation in market research. For example, we have used latent-class DCE models for HIV screening in Tanzania. In another example, a latent-class DCE of colorectal cancer screening, we found three groups of patients: those concerned about accuracy, another about cost and another about discomfort.81 We will use a main effect mixed logit model and models with interaction terms in Aim 2 between DCE attributes and individual-specific factors. The mixed logit model will be used to estimate the mean and SD of attribute utility, where higher utility indicates greater relative utility (ie, more preferred attributes compared with others included in the DCE). Stata V.16 will be used for analysis.82

Patient and public involvement

While the research questions were developed by the study team, the outcomes (attributes) for the DCE will be derived from the qualitative data collected from research participants. Additionally, the eventual preferences will be established by participants through the DCE. No patients were involved in this study. Results from the DCE will be shared back with the community through established Community Advisory Boards.