Promoting HIV Testing by Men: A Discrete Choice Experiment to Elicit Preferences and Predict Uptake of Community-Based Testing in Uganda

Abstract

Background & objectives:

HIV testing is essential to access HIV treatment and care and plays a critical role in preventing transmission, yet testing coverage is low among men in sub-Saharan Africa. Community-based testing has demonstrated potential to expand male testing coverage, yet scant evidence reveals how community-based services can be designed to optimize testing uptake. We conducted a discrete choice experiment (DCE) to elicit preferences and predict uptake of community-based testing by men in Uganda.

Methods:

Hypothetical choices between alternative community-based testing services and the option to opt-out of testing were presented to a random, population-based sample of 203 adult male residents. The testing alternatives varied by service delivery model (community health campaign, counselor-administered home-based testing, distribution of HIV self-test kits at local pharmacies), availability of multi-disease testing, access to antiretroviral therapy (ART), and provision of a US$ 0.85 incentive. We estimated preferences using a random parameters logit model and explored whether preferences varied by participant characteristics through subgroup analyses. We simulated uptake when a single and when two community-based testing services are made available, using reference values of observed uptake to calibrate predictions.

Results:

The share of the adult male population predicted to test for HIV ranged from 0.15 to 0.91 when a single community-based testing service is made available and from 0.50 to 0.96 when two community-based services are provided concurrently. ART access was the strongest driver of choices (relative importance [RI]=3.01, 95% confidence interval [CI]: 1.74–4.29), followed by the service delivery model (RI=1.27, 95% CI: 0.72–1.82) and availability of multi-disease testing (RI=1.27, 95% CI: 0.09–2.45). A US$ 0.85 incentive had the least yet still significant influence on choices (RI=0.77, 95% CI: 0.06–1.49). Men who perceived their risk of having HIV to be relatively elevated had higher predicted uptake of HIV self-test kits at local pharmacies, as did young adult men compared to men aged ≥30 years. Men who earned the daily median income or less had higher predicted uptake of all community-based testing services than men who earned above the daily median income.

Conclusion:

Substantial opportunity exists to optimize the delivery of HIV testing services to expand uptake by men; using an innovative DCE, we deliver timely, actionable guidance for promoting community-based testing by men in Uganda. We advance the stated preference literature methodologically by describing how we constructed and evaluated a pragmatic experimental design, used interaction terms to conduct subgroup analyses, and harnessed participant-specific preference estimates to predict and calibrate testing uptake.

1. INTRODUCTION

The Joint United Nations Program on HIV and AIDS has proposed a set of ambitious objectives known as the 90–90-90 targets that countries should attain by 2020 in order to end the AIDS epidemic as a major global health threat by 2030. Increasing awareness of HIV status such that 90% of HIV-positive persons know their status constitutes the first target and sets the standard for progress toward the remaining two targets [1]. While considerable progress has been made to expand HIV testing access, a number of important gaps in testing coverage exist, including low uptake by men in sub-Saharan Africa. Men are less likely than women to test for HIV and know that they are HIV-positive [2,3]. Consequences of the testing disparity include missed or late HIV diagnosis, delayed initiation of antiretroviral therapy (ART), and increased mortality for HIV-positive men [4–9]. Low rates of male testing also constitute missed opportunities to prevent transmission as HIV testing provides not only an entry point for linking persons to care for their own health but for providing treatment, counseling, and additional interventions for HIV prevention [10,11].

Discovering how to optimize the delivery of HIV testing to expand uptake by men is a vital need for decision makers who seek to attain the 90–90-90 targets. Community-based HIV testing is an approach that has achieved higher coverage among men than standard facility-based testing [12,13]. Community-based service delivery models offer enhanced convenience and accessibility and provide flexible platforms to administer interventions to promote testing uptake. Community-based service delivery models that have been widely implemented in sub-Saharan Africa include mobile, home, workplace and event-based service delivery models. Additionally, the World Health Organization (WHO) has called for the expansion of service delivery models that use oral fluid-based HIV self-tests as novel technologies to promote testing, particularly among hard-to-reach and at-risk populations [14]. Possibilities to leverage community-based services to promote testing are therefore abundant, yet scant evidence reveals particular service delivery models and interventions that will induce the greatest uptake by men.

This paper presents results from a discrete choice experiment (DCE) to elicit men’s preferences and predict uptake of community-based HIV testing in rural Uganda. A DCE yields preference estimates that reveal the relative influence of service attributes on hypothetical choices between alternative services that participants make; the preference estimates can also be used predict service uptake [15,16]. Uptake predictions are meaningful for decision-makers as they reveal population shares that are expected to use particular services under conditions specified in the DCE. Further, uptake predictions can be incorporated into mathematical models to forecast health outcomes and the cost-effectiveness of alternative strategies to promote testing. Given the hypothetical nature of a DCE, uptake predictions for services that have and have not been previously implemented can be investigated. DCEs have been employed in marketing, transportation, and natural resource economics, and grown increasingly prominent in health economics. Several studies have investigated preferences for HIV testing, including among diverse populations in sub-Saharan Africa [17–21], yet none have used a DCE to predict testing uptake.

Our use of a DCE to elicit men’s preferences and predict HIV testing uptake in Uganda is highly innovative and informative. We estimate preferences for service delivery models that hold strong potential to appeal to men and for additional attributes of community-based testing that represent timely interventions that could be undertaken to increase male testing. We then simulate testing uptake under implementation scenarios of interest to decision makers who seek to expand male testing coverage. Our findings are important because they reveal the comparative potential of promising service delivery models and interventions to promote male testing. We also advance the stated preference literature methodologically by describing how we constructed and evaluated a pragmatic experimental design and employed a hybrid coding scheme that ensured that attribute variables were not confounded with a model constant for opting-out of HIV testing. We further detail how we investigated preference heterogeneity by participant subgroups, using subgroup interactions that overcome the challenge of scale differences that arises when estimating separate regressions for each subgroup. We also describe how we harnessed participant-specific preference estimate to predict testing uptake and calibrated our uptake predictions using observed estimates of testing uptake for the study setting. Each of these contributions helps advance the methods base for applying stated preference studies to inform the optimal design and delivery of health services.

2. METHODS

2.1. Setting and sample

This study was conducted in Mbarara District, a rural region of southwestern Uganda where HIV prevalence among adult males is 7.0 – 8.0% [22]. The DCE was administered during the enrollment phase (April – June 2016) of a randomized trial investigating the comparative effectiveness of novel incentive strategies to encourage men to test for HIV at a local, multi-disease community health campaign (NCT02890459). Prior to enrollment for the trial, a household census was conducted in four neighboring parishes of Mbarara District. The parishes fell within the catchment area of one PEPFAR-supported government clinic. Although a number of community-based HIV testing strategies have been implemented in Uganda in recent years, experience with community-based testing is limited and facility-based testing is standard. Male residents who were ≥18 years of age were eligible to participate in the trial if they had been living in the community for at least six months in the past year and were not planning to move away in the next three months. All eligible men present at the time of the census were invited to participate in the trial.¹ Men who agreed to participate were administered an enrollment questionnaire. The DCE formed one section of the questionnaire.

2.2. Attribute development

Attributes of HIV testing that can be leveraged to promote testing were developed based on a literature review. We first reviewed the literature to identify service delivery models that have been implemented or are under active consideration in sub-Saharan Africa and selected three based on their potential to increase HIV testing by men in Uganda: 1) HIV testing at a community health campaign; 2) counselor-administered home-based testing; and 3) distribution of oral fluid-based HIV self-test (HIVST) kits at local pharmacies. We then added attributes to the DCE to represent changes to how community-based testing is delivered (i.e. interventions) that could promote testing. We included three binary attributes that indicated whether: 1) multi-disease testing; 2) immediate access to antiretroviral therapy (ART) for HIV-positive persons; and 3) a financial incentive were available. The service delivery model and intervention attributes are described below, and the levels used to describe the attributes to participants are presented in Table 1.

Table 1.

DCE attributes and levels

Attribute		Levels
Service delivery model	-	You attend a community health event in your village
	-	A health counsellor comes to your home and offers to test you for HIV
	-	You pick up a self-test kit at a nearby pharmacy
Availability of multi-disease testing services	-	Only HIV testing is available
	-	You can test for tuberculosis, malaria, pressure and diabetes when you test for HIVa
Access to ART for HIV-positive persons	-	Medications to treat HIV are not immediately available
	-	Medications to treat HIV are immediately available
Provision of a testing incentive	-	You do not receive compensation
	-	You receive 3,000 Shillings when you test for HIV

ART, antiretroviral therapy

^a“Pressure” was used to refer to hypertension/high blood pressure.

2.2.1. Community-based service delivery models

HIV testing at a community health campaign:

Community health campaigns are a form of mobile testing that have achieved high population-level testing coverage in several parts of sub-Saharan Africa [23–26]. Community health campaigns are held at convenient locations and typically offer HIV testing with other health services. HIV testing uptake by men at prior community health campaigns is high relative to facility-based testing [23].

Counselor-administered home-based testing:

Through home-based testing, a health counselor makes door-to-door visits to households in a community and offers to test household members for HIV. Home-based testing has been implemented in Uganda and found to be effective at reaching population groups with low rates of prior testing [27,28]. Previous findings suggest men prefer home-based testing to facility-based testing [29,30].

Distribution of HIVST kits at local pharmacies:

Oral fluid-based HIVST kits allow users to take their own sample, perform a simple test, and interpret the result on their own. HIVST has been shown to have high acceptability for diverse populations, including groups that are less likely to access facility-based testing [31–33]. HIVST kits are not yet widely available in Uganda, and we included the distribution of HIVST kits at local pharmacies as a service delivery model that could be introduced.²

2.2.2. Interventions to promote testing

Availability of multi-disease testing:

Providing testing for multiple diseases at the time of HIV testing could reduce barriers due to HIV-related stigma and appeal to men who perceive their risk of being HIV-positive to be low. Integration of testing services has been shown to be feasible in several countries, including Uganda, and findings indicate that persons who access community-based HIV testing that is integrated with testing for other diseases are more likely to be male than those who access facility-based HIV testing [23,34]. We included the ability to test for tuberculosis, malaria, hypertension, and diabetes at the time of HIV testing as an attribute in the DCE.³

Access to ART for HIV-positive persons:

Ensuring immediate ART access for persons who test HIV-positive could be reassuring to men who are uncertain of their status. Findings from multiple qualitative studies suggest that ensuring ART access following HIV diagnosis is likely to promote testing and may help men overcome social barriers to testing [35–38]. Many countries including Uganda have recently adopted WHO’s “Treat All” guideline recommendation that all individuals who are diagnosed with HIV should start ART, regardless of disease stage or CD4 count [11,39]. Practice guidelines therefore now uphold ART initiation directly following HIV diagnosis, and it is highly timely to assess the impact that providing immediate ART access for HIV-positive persons could have on HIV testing uptake.

Financial incentive for HIV testing:

Offering incentives has been proposed as an intervention to increase male testing and could help offset financial or psychosocial costs that men associate with testing [40–43]. A systematic review found that incentives are effective at increasing HIV testing for diverse populations, especially when testing is provided outside of health facilities [40]. We included the provision of a fixed incentive of 3,000 Ugandan shillings (about US$ 0.85) for HIV testing as an attribute in the DCE. This amount was less than the median daily income of adult male residents of the study district and can be considered a relatively small sum that could help offset costs from time spent not working or for transportation to a local testing venue.

2.3. Choice set format

Choice sets were constructed to display two alternatives for community-based HIV testing and an opt-out alternative so that participants could choose not to test if neither testing alternative appealed to them. Each testing alternative was defined by a set of attribute levels. When pairing a specific level of the service delivery model attribute with certain levels of the intervention attributes would result in testing alternatives that policymakers would not consider implementing, we defined constraints such that the attribute levels were not paired (Appendix A).

2.4. Experimental design construction

The fundamental task of experimental design construction is to distribute attribute levels across choice set alternatives such that a design is able to determine the independent effect that each attribute has on participant choices and detect statistically significant relationships between the attributes and participant choices. For this study’s purposes, prediction of testing uptake was a principal objective and design constraints were important to ensure that all testing alternatives presented in the design were alternatives that policymakers would consider implementing. Additionally, integrating the DCE into the enrollment phase of a randomized trial provided a unique opportunity to administer the DCE to a population-based sample yet imposed practical constraints. The trial required a large sample, and time available to administer the DCE was limited. Administration of the DCE was restricted to a subsample of trial participants; a limited number of choice set questions could also be presented. Given our research objectives and constraints, we prioritized design efficiency over strict orthogonality.

After imposing design constraints, ten testing alternatives were possible that could be combined in 45 unique choice sets, and we generated a fractional factorial design of 10 carefully selected choice sets. Although we had access to experimental design software, the software did not incorporate prior information regarding preferences. Prior evidence suggested that preferences for the levels of the 3 intervention attributes would be ordered such that most participants would prefer that: 1) testing services be available for multiple diseases (including HIV) rather than for HIV alone; 2) access to ART for HIV-positive persons be immediately available rather than not immediately available; and 3) a US$ 0.85 testing incentive be provided rather than not provided. The preference orderings for the intervention attribute levels were a unique feature of our design that could be used to increase the design efficiency. We realized that we could enhance the design efficiency if we manually constructed our experimental design and chose to do so.

We enumerated all 45 possible choice sets and first excluded those that included a dominant testing alternative (n=11). Dominant testing alternatives were identified in choice sets where the level of the service delivery model attribute was the same across testing alternatives as we made no assumptions regarding preferences for the service delivery models presented in the DCE. We then evaluated the remaining choice sets according to the trade-offs between attributes that each choice set represented. Trade-offs were assessed as differences in attribute levels across testing alternatives. We considered the number of trade-offs represented and the directionality of the trade-offs for differences between the binary attribute levels. We selected 10 choice sets that represented a wide range of trade-offs that participant could make between testing alternatives. Most of the selected choice sets (60%) included relatively complex trade-offs between three or four attributes to promote statistical efficiency, yet we also selected several choice sets (40%) that included simpler trade-offs between one or two attributes to promote response efficiency by limiting measurement error [44]. As a final step, we used random assignment to block the experimental design into 2 versions of 5 choice sets.

2.5. Experimental design evaluation

Evaluation of the correlation properties of our experimental design was imperative given our manual design construction. We evaluated the correlation properties of our experimental design by calculating J-index correlation coefficients [45,46]. We calculated J-index correlation coefficients to assess correlation of the differences in attribute levels across testing alternatives in the experimental design, in accordance with how logit and probit-based random utility models are derived. J-index correlation coefficients range from 0 to 1, and values closer to 1 indicate stronger correlation. The J-index correlation coefficients indicated weak correlations of the differences in attribute levels across testing alternatives in our experimental design, with values that ranged from 0.03 to 0.26 (Appendix B).

We applied a widely used rule proposed by Johnson and Orme (2010) to assess the sample size needed for our experimental design. The rule recommends that the sample size (N) to estimate main effects satisfy the following equation:

$$N\ge \frac{500c}{(t\times a)}$$

where c is the largest number of levels for any one attribute (i.e. analysis cells), t is the number of choice sets (i.e. tasks) and a is the number of alternatives in each choice set, excluding the opt-out alternative [47]. By this rule, we needed a sample size of at least 150 participants, and our projected sample size of 200 DCE participants was satisfactory.

Our primary evaluation of the experimental design indicated that the design was adequate to estimate the independent effects of the design attributes on participant choices. Our evaluation was based on a model specification that included only attribute variables (i.e. main effects) yet we questioned whether the effects of the intervention attributes could vary by service delivery model. Assessing our design for inclusion of interaction terms between the intervention and service delivery model attributes indicated that a larger sample size or more choice sets were warranted to ensure that the statistical significance of the interaction terms could be detected at conventional thresholds. Additionally, J-index coefficients calculated to assess correlation of the differences in the levels of the attribute variables and interactions across testing alternatives were mostly unconcerning yet revealed two strong, potentially problematic correlations. As we had no theory-driven hypotheses that motivated the inclusion of attribute interaction terms and needed to respect research environment constraints, we proceeded with the design here described.

2.6. Survey administration and piloting

Twenty-four enumerators were trained to enroll participants into the randomized trial and administer the enrollment questionnaire (Appendix C). The enumerators used handheld tablets to administer the questionnaire, and the DCE was programmed for random delivery to 1 in 10 participants. Before beginning the DCE, the enumerators read a standardized script that introduced the experiment and provided information about community-based HIV testing as an approach that is distinct from facility-based testing. Participants were asked to imagine that the testing alternatives presented in the DCE were going to be made available in their communities in the next 12 months. The introductory script and DCE choice sets were translated from English into Runyankole, the local language, and read aloud. Reading the choice sets aloud helped ensure that participants understood the choice sets despite varying levels of literacy. Additionally, the enumerators were provided paper booklets that contained illustrations of the choice sets, with simple graphics used to represent each attribute level, and the enumerators used the booklets to describe the choice sets to participants. Fig. 1 presents an example choice set.

Fig. 1.

Example choice set

During the first two weeks of enrollment into the randomized trial, we piloted a version of the DCE. Piloting allowed the enumerators to practice and hone their delivery of the DCE. The principal investigator of the DCE accompanied the enumerators during the piloting phase to ensure that the DCE was administered consistently. She provided coaching to individual enumerators and debriefed with the team of enumerators at the end of each day to collectively address questions that arose during the day. Piloting also allowed us to assess participant understanding of the choice sets. We revised the wording of the attribute levels and adjusted the formatting of the choice sets to enhance comprehension. We further examined response data to confirm that participants were attending to the choice sets and were not, for instance, always choosing the same response option to expedite progression through the questionnaire.

2.7. Conceptual framework and econometric modeling

Random utility theory provides the conceptual framework for discrete choice analysis. According to random utility theory, the utility U_nit a participant n derives from alternative i in choice set t consists of a deterministic component V_nit and an unobservable random component ε_nit such that:

$$U_{nit}=V_{nit}+{\epsilon }_{nit}={{\beta }^{\prime }}_n{x}_{it}+{\epsilon }_{nit}$$ (1)

where x_it is a vector of observed variables (i.e. attributes) that describe alternative i and β_n is a vector of marginal utility (i.e. preference) parameters associated with the variables. It is assumed that participant n makes choices to maximize his utility such that the probability that he chooses alternative i over another alternative j in choice set t is expressed as:

$${\mathrm{Pr}}_{nit}=\mathrm{Pr}(U_{nit}>U_{njt})=\mathrm{Pr}(V_{nit}+{\epsilon }_{nit}>V_{njt}+{\epsilon }_{njt})=\mathrm{Pr}(V_{nit}-V_{njt}>{\epsilon }_{njt}-{\epsilon }_{nit})$$ (2)

Assuming that ε_nit are independently and identically distributed (iid) over alternatives, people, and choices leads to a multinomial logit (MNL) model. Conditional on β_n, the probability that participant n chooses alternative i in choice set t is provided by:

$$L_{nit}({\beta }_n)=\frac{\exp ({{\beta }^{\prime }}_n{x}_{it})}{{\sum }_{j=1}^J\exp ({{\beta }^{\prime }}_n{x}_{jt})}$$ (3)

and the probability that participant n makes a particular sequence of choices is a product function:

$$S_n({\beta }_n)={\prod }_{t=1}^T{L}_{ni(n,t)t}({\beta }_n)$$ (4)

where i(n,t) is the alternative chosen by participant n in choice set t [48]. Although β_n cannot be observed, choices can be and β_n can easily be estimated given the assumption that ε_nit are iid over alternatives, people, and choices. The assumption is strong, however, and disallows unobserved preference heterogeneity, correlation in unobserved factors, and unrestricted substitution patterns. More flexible models have been advanced whose strengths can be understood with regard to how they overcome the limitations of the MNL model [49–54].

We employed a random parameters logit (RPL) model to estimate preference parameters. The RPL model allows preference parameters to vary over participants. The vector of parameter estimates β_n can be expressed as the sum of the population mean b and the person-specific deviation η_n, which represents the preferences of participant n relative to the population mean [55]:

$$U_{nit}={{\beta }^{\prime }}_n{x}_{it}+{\epsilon }_{nit}=(b^{\prime }+{{\eta }^{\prime }}_n)x_{it}+{\epsilon }_{nit}$$ (5)

The RPL model can also be understood as a specification that decomposes the random utility component from Eq. 1 into two terms $({{\mathit{\eta }}^{\prime }}_{\mathit{n}}{\mathit{x}}_{\mathit{i}\mathit{t}}+{\mathit{\epsilon }}_{\mathit{n}\mathit{i}\mathit{t}})$ [52]. Correlation over alternatives and choices is allowed through ${\mathit{\eta }}_{\mathit{n}}$, and the remaining term, ${\mathit{\epsilon }}_{\mathit{n}\mathit{i}\mathit{t}}$ represents the unobserved portion of utility assumed to be iid over alternatives, people, and choices.

The RPL model is an appealing specification for our choice context as it is likely that preference heterogeneity exists across our population-based sample and that choices made by the same participant are correlated. The RPL model also holds important prediction benefits that can be perceived from the unconditional and simulated modeling of choice probabilities. The probability that participant n makes a particular sequence of choices is provided by the integral of the multinomial probability (Eq. 4) over all possible values of β:

$$P_n(\theta )=\int S_n(\beta )f(\beta |\theta )d\beta$$ (6)

where $\mathit{f}(\mathit{\beta }|\mathit{\theta })$ is the density of β. The researcher specifies the distribution of β and the goal is to estimate θ, the parameters that describe the distribution (e.g. often the mean and variance for continuous distributions). Although the integral cannot be solved analytically, it can be estimated through maximum likelihood simulation, and the simulated log-likelihood is given by:

$$SL{L}_{RPL}(\theta )={\sum }_{n=1}^N\ln \left[\frac{1}{R}{\sum }_{r=1}^R{S}_n({\beta }^r)\right]$$ (7)

where R is the number of replications and β^r is the rth draw from $\mathit{f}(\mathit{\beta }|\mathit{\theta })$.

From Eq. 6, it can be seen that the ratio of any two choice probabilities necessarily depends on all of the data—the specification of attribute variables and the distribution of β [56]. Correlation between choice probabilities is accounted for in the simulation of Eq. 7 and substitution patterns between alternatives are thus highly flexible. This flexibility is essential to generate accurate choice predictions. Further, in addition to allowing preference parameters to vary over participants, the RPL model can be used to estimate participant-specific preferences. Participant-specific preference parameters can be derived as the expected value of β, conditional on a given response pattern y_n and set of variables x_it that describe all alternatives [57]:

$$E\left[\beta |y_n,x_{it}\right]=\frac{\int \beta S_n(\beta )f(\beta |\theta )d\beta }{\int S_n(\beta )f(\beta |\theta )d\beta }$$ (8)

Estimation of participant-specific parameters ${\widehat{\mathit{\beta }}}_{\mathit{n}}$ is also achieved through simulation:

$${\widehat{\beta }}_n=\frac{\left(\frac{1}{R}\right){\sum }_{r=1}^R{\beta }^r{S}_n({\beta }^r)}{\left(\frac{1}{R}\right){\sum }_{r=1}^R{S}_n({\beta }^r)}$$ (9)

where ${\mathit{\beta }}^{\mathit{r}}$ is the rth draw from $\mathit{f}(\mathit{\beta }|\widehat{\mathit{\theta }})$ [57]. Although many applications of the RPL model conclude upon estimation of θ (Eq. 7), the additional step of estimating participant-specific preferences has been demonstrated to considerably improve predictions [57].

2.8. Model specification, hybrid coding, and estimation

To estimate men’s preferences for attributes of community-based HIV testing in rural Uganda, we assumed a linear-in-parameters specification represented as:

$$\begin{gathered} V_{nit}-V_{nCBTest}={\beta }_{nCHC}Communityhealthcampaig{n}_{it}+{\beta }_{nHBT}Home\text{\_}basedtestin{g}_{it}+ \\ {\beta }_{nMDT}Multi\text{\_}diseasetestin{g}_{it}+{\beta }_{nART}ImmediateAR{T}_{it}+{\beta }_{nInc}Incentiv{e}_{it}+ \\ {\beta }_{nWNT}Wouldnottes{t}_{it} \end{gathered}$$ (10)

where Community health campaign_it, Home_based testing_it, Multi_diesease testing_it, Immediate ART_it, and Incentive_it represent attribute variables that described the testing alternatives and Would not test_it is an alternative-specific constant that accounts for the fact that a participant could always choose to opt-out of HIV testing, given the testing alternatives presented.

A categorical data coding scheme was required as the attribute levels were primarily qualitative. Dummy and effects coding schemes are commonly employed for the analysis of discrete choice data. Dummy coding uses a series of 0s and 1s to relate each level of an attribute to an omitted base level for the attribute that serves as a referent and is set at zero. A dummy coding scheme results in a base level for each attribute and the coefficients for the levels of one attribute cannot be compared to coefficients for the levels of another. The base levels are also confounded with a model constant, if one is included. Effects coding instead uses a series of 0s, 1s, and −1s to center all attribute coefficients on a grand mean of zero. The omitted level for each attribute is coded as −1 and its coefficient is retrieved as the negative sum of the coefficients for the levels of the attribute included in estimation. An effects coding scheme allows all attribute level coefficients to be interpreted against a common referent and ensures that the attribute level coefficients are not confounded with a model constant [58]. Yet, if the levels of an attribute do not pertain to a particular alternative (e.g. in the case of an opt-out alternative), coding the attribute variables with effects codes for the alternative leads to confounding with the model constant, and a hybrid coding scheme can overcome this challenge [59].

We employed a hybrid coding scheme in which we effects coded the model variables for all attribute levels that pertained to the testing alternatives and coded the variables as zero (i.e. dummy coded) for the opt-out alternative, given that none of the attribute levels pertained to the opt-out alternative. The opt-out constant itself was also dummy coded. This hybrid coding scheme achieved a particular, useful interpretation of model coefficients. The coefficient for the single model constant for the opt-out alternative was interpreted as the utility associated with choosing to opt-out of HIV testing relative to choosing to test for HIV, given the testing alternatives presented in the DCE. The effects-coded attribute level coefficients were interpreted as marginal utility departures that each attribute level rendered to the testing alternatives, conditional on HIV testing being chosen (i.e. the omitted base level for the dummy-coded constant). The design grand mean was thus centered on choosing to test for HIV and could be considered the average utility that a participant assigned to testing (V_nCBTest), given the testing alternatives presented in the DCE.

The main effects utility function provided in Eq. 7 was estimated using a RPL model with random normal distributions for all model parameters. We compared the RPL model to a MNL model to confirm that the RPL model improved model fit and investigated whether model fit was further improved by using lognormal distributions for the intervention attribute parameters or by adding interactions between the service delivery model and intervention attributes to the model specification. Model fit was evaluated according to log-likelihood values, Akaike information criteria (AIC), and likelihood ratio (LR) test statistics. Because our calculation of J-index correlation coefficients indicated that differences in the levels of certain attribute variables and attribute interaction terms were strongly correlated, we interrogated for multicollinearity when including interactions in the model specification by estimating variance inflation factors (VIFs). Multicollinearity is typically investigated when analyzing revealed rather than stated preference data, yet our non-orthogonal experimental design allowed some correlation among variables and interaction terms, and we investigated whether the correlation was strong enough that it could be detected as multicollinearity. Analyses were conducted using Stata version 15 (StataCorp, College Station, TX).

2.9. Subgroup analyses

The primary results from a RPL model include mean preference estimates and standard deviations that indicate the extent to which preferences vary in the sample. Decision-makers who seek to increase male testing may be interested to know how preferences vary according to participant characteristics, and we used data collected via the enrollment questionnaire to conduct subgroup analyses. We generated indicator variables to distinguish subgroups by self-reported characteristics, including HIV testing history, perceived risk of having HIV, and sexual behaviors in the 12 months prior to enrollment. We also generated indicators to distinguish subgroups demographically by age, marital status, and income.⁴ For each subgroup analysis, we interacted the indicator variable of interest with the attribute variables, attribute interactions, and opt-out constant. We then estimated the fully interacted model with fixed parameters for the subgroup interactions and random normal distributions for all other parameters. Using interactions to conduct subgroup analyses in this manner is equivalent to estimating separate regressions per subgroup yet avoids the need to account for scale differences that arises when estimating separate regressions [60]. The subgroup analyses were exploratory as the DCE was not powered to detect the significance of the subgroup interactions at conventional thresholds, and we instead investigated whether including subgroup interactions significantly improved model fit by conducting LR tests.

2.10. Uptake simulation and calibration

Predictions of HIV testing uptake were generated using participant-specific preference coefficients. After obtaining the mean preference coefficients and standard deviations for the sample by simulating the log-likelihood function provided in Eq. 7, we simulated participant-specific preferences according to Eq. 9. Using the participant-specific coefficients, we calculated the utility that each participant assigned to the testing alternatives that were presented in the DCE and to the opt-out alternative. We then applied a utility maximization rule that a participant would choose the alternative in a given scenario that provided him the greatest utility (i.e. the assumption upon which DCEs are based) and simulated testing uptake when: 1) a single community-based testing service is made available and 2) two community-based testing services are made available. Decision makers who seek to promote male testing are likely first interested to know predicted uptake when a single community-based testing service is implemented. Yet, to attain the ambitious 90–90-90 targets, decision makers may consider implementing two testing services in tandem.

To ensure that simulated uptake predictions were consistent with reference values of observed uptake, we implemented a calibration procedure described by Train (2009) [56]. The procedure involves iteratively adjusting the alternative-specific constant(s) which captures the average effect of factors that are not included in the model on utility for an alternative using the following formula:

$${\alpha }_j^1={\alpha }_j^0+\ln (S_j/S_j^0)$$ (11)

where ${\alpha }_j^0$ is the constant estimated by the discrete choice model. S_j is the share of the population (i.e. adult male residents) who have been observed to choose alternative j in the forecast environment (i.e. rural Uganda), and $S_j^0$ is the share of participants who are predicted to choose alternative j using DCE-derived preference estimates [56]. The new constant ${\alpha }_j^1$ is used to predict uptake and is adjusted again as needed until the share of participants predicted to choose alternative j matches the observed population share who chose alternative j. The calibration constant is then used to make all other predictions (i.e. for scenarios where testing uptake by the population has not been previously observed). The only alternative-specific constant that we included in our model was for the opt-out alternative which captured the effect of unobserved factors related to choosing to opt-out versus use community-based HIV testing. A key assumption that we made was that unobserved factors influenced choices to opt-out versus opt-in to community-based testing yet did not influence choices between testing alternatives.

We performed the calibration procedure at the participant level, iteratively adjusting the participant-specific opt-out constants estimated by the RPL model until the predicted testing uptake for the sample matched observed testing uptake as reported by Chamie et al. (2016). Chamie et al. report HIV testing uptake by adult residents of rural communities in Uganda and Kenya following implementation of a hybrid testing strategy [28]. Residents were first invited to participate in a community health campaign where HIV testing was provided with testing for other diseases (i.e. multi-disease testing); immediate access to ART and financial incentives were not provided. Residents who did not attend the campaign were then offered counselor-administered home-based testing in which HIV testing alone was provided (i.e. multi-disease testing was not provided); immediate access to ART and financial incentives were not provided. The hybrid testing strategy thus provided reference values for observed testing uptake in a setting very similar to our study setting when both a single testing service was made available and when two testing services were implemented in rapid succession.

3. RESULTS

3.1. Sample characteristics

In total, 91% of eligible adult male residents agreed to participate and were enrolled in the randomized trial. The majority (67%) of trial participants were enrolled after the DCE pilot-testing phase, of whom 203 participants (12%) were randomly selected to receive the DCE. All trial participants who were randomized to receive the DCE completed the DCE. Nine participants who completed the DCE reported an HIV-positive status at enrollment and were excluded from analysis. Characteristics of the 194 participants who comprised the analytical sample are presented in Table 2. For comparison, the DCE sample characteristics are presented alongside characteristics of adult male residents of Uganda as assessed by the 2016 Demographic and Health Survey, a nationally representative data source [61].

Table 2.

Sample characteristics compared to nationally representative data^a

	DCE sampleb	Rural adult male residents	P-value of test statisticc
Total participants, N	194	3,502
Age (years), median (IQR)	34.5 (26–48)	30 (23–40)	<0.01
Age group (years)
18–29	34%	48%	<0.01
30–49	42%	47%
50+	24%	6%
Highest level of education attained
Primary or less	70%	64%	0.20
Secondary	23%	25%
Beyond secondary	7%	10%
Occupation
Agriculture/farming	52%	54%	0.48
Manual labor	22%	20%
Professional/business/sales/services	17%	19%
Transportation/other	10%	7%
Marital status
Married or cohabitating	63%	68%	0.11
Never married	26%	25%
Separated/divorced/widowed	11%	7%
Tested for HIV in the past 12 mo.
No	47%	42%	0.20
Yes	53%	58%
Number of sexual partners in the past 12 mo.
No partners	17%	14%	0.68
1 partner	61%	62%
2 or more partners	23%	24%
Perceived risk of having HIVd
No risk	34%	--	--
Low risk	45%
Moderate or high risk	18%
Unknown risk	3%

DCE, discrete choice experiment; IQR, interquartile range.

^a2016 Demographic and Health Survey (DHS) data were used to compare sample characteristics with characteristics of adult (aged ≥18 years) male residents of Uganda. DHS sampling weights have been applied.

^bExcludes 9 participants who reported that they were HIV-positive at enrollment. HIV serostatus is not assessed by DHS; including the 9 participants in the DCE sample does not considerably alter comparisons of sample characteristics with nationally representative survey data.

^cP-values are for χ2 test statistics of equal sample proportions; a nonparametric test corrected for continuity was used to compare median age across samples.

^dPerceived risk of having HIV is not assessed by DHS.

The median age of DCE participants was 34.5 years (interquartile range: 26–48 years). Most participants had a primary education or less (70%), were employed in agriculture or manual labor (74%), and were married (63%). Over half (53%) reported that they tested for HIV in the past 12 months, 95% of whom reported testing at a health facility. Twenty-three percent of participants reported having two or more sexual partners in the past 12 months, and 18% perceived their risk of having HIV to be moderate or high. Comparison of sample characteristics with DHS data revealed that DCE participants were on average older yet were otherwise representative of adult male residents of rural Uganda.

3.2. Sample preference estimation

Estimation of MNL and RPL models confirmed that the RPL model dramatically improved model fit compared to the MNL model, regardless of whether models without interactions (LR test statistic=1130.91, p<0.01) or with interactions that allowed the effect of the intervention attributes to vary by service delivery model (LR test statistic=1140.39, p<0.01) were compared (Appendix D). Proceeding with the RPL model, we found that model fit was not improved by using lognormal distributions (singly or jointly) for the intervention attribute parameters yet was improved by adding attribute interactions to the model specification (LR test statistic=22.13, p<0.01). VIFs estimated after running the RPL model with attribute interactions did not indicate that multicollinearity was problematic. We selected the RPL model with attribute interactions as our final specification. The results of the RPL with attribute interactions are presented in Table 3, and we discuss the remainder of our results with respect to this specification.

Table 3.

Preference estimation

	β	SE	SD	SE
Alternative-specific constant
Opt-out of HIV testing	−9.59***	(2.62)	4.67**	(1.54)
Main effects
Service delivery model
Community health campaign (CHC)	0.13	(0.38)	1.14*	(0.57)
Home-based testing (HBT)	0.57*	(0.22)	0.37	(0.62)
HIV self-testing (HIVST)a	−0.70**	(0.25)	1.50**	(0.48)
Availability of multi-disease testing
Only HIV testing is availablea	−0.63*	(0.30)	0.01	(0.14)
Multi-disease testing is available	0.63*	(0.30)	0.01	(0.14)
Access to antiretroviral therapy
ART is not immediately availablea	−1.51***	(0.33)	1.53***	(0.34)
ART is immediately available (ART)	1.51***	(0.33)	1.53***	(0.34)
Provision of an incentive
No incentivea	−0.39*	(0.18)	0.75	(0.39)
Incentive of US$ 0.85 (Inc)	0.39*	(0.18)	0.75	(0.39)
Attribute interactions
CHC*ART	0.58	(0.31)	0.61	(0.41)
HBT*ARTa	−0.58	(0.31)	0.61	(0.41)
CHC*Inc	0.04	(0.35)	1.82***	(0.53)
HIVST*Inca	−0.04	(0.35)	1.82***	(0.53)
Goodness of fit statistics
Log-likelihood	−523.96
AIC	1079.92
Likelihood ratio tests
Test statistic
Test statistic	1140.39
P-value	<0.01
RPL with vs. RPL without interactions
Test statistic	22.13
P-value	<0.01
No. of observations (choices)	970

β, Beta coefficient; SE, standard error; SD, standard deviation; ART, antiretroviral therapy; US$, United States dollars; AIC, Akaike information criteria; RPL, random parameters logit; MNL, multinomial logit. Results are reported as mean preference and standard deviation coefficients, with robust standard errors clustered at the individual level in parentheses; RPL models were simulated with 500 Halton draws;

^***p<0.001,

^**p<0.01,

^*p<0.05.

^aThis attribute level was omitted during estimation. Coefficients for this level and their standard errors were retrieved post-estimation.

The strong negative and statistically significant preference coefficient for the opt-out constant indicates that participants largely preferred to choose one of the testing alternatives presented in the choice sets rather than the opt-out alternative. Preference coefficients for the attribute variables were mostly significant, indicating that the attribute levels included in the DCE influenced choices between testing alternatives, conditional on HIV testing being chosen. The rank ordering of coefficients for the service delivery model attribute reveal that participants generally preferred counselor-administered home-based testing, followed by HIV testing at community health campaign, and distribution of HIVST kits at local pharmacies. The only difference found to be statistically significant, however, was between coefficients for counselor-administered home-based testing and distribution of HIVST kits at local pharmacies (p<0.01). The coefficients for the intervention attributes indicate that participants preferred HIV testing be provided along with testing for other diseases rather than alone (p=0.04), that ART be immediately available at the time of testing rather than not immediately available (p<0.01), and that a US$ 0.85 incentive be provided rather than not provided (p=0.04).

The magnitude of the preference coefficients for the attribute levels can be compared across attributes given the effects coding of the attribute variables. To facilitate interpretation of the importance of each attribute to participants relative to the other attributes included in the DCE, we graphed the attribute relative importance (RI) in Fig. 2. The RI was assessed as the greatest difference between coefficients associated with the levels of an attribute (i.e. the difference between the most positive and the most negative coefficients); this difference represents the maximum utility change that can be achieved from an attribute, conditional on the range of attribute levels included in the DCE. The attribute with the greatest RI was access to ART at the time of HIV testing (RI=3.01, 95% confidence interval [CI]: 1.74–4.29). The RI of the service delivery model attribute ranked second (RI=1.27, 95% CI: 0.72–1.82) and was equivalent to the RI of the availability of multi-disease testing attribute (RI=1.27, 95% CI: 0.09–2.45). Provision of a testing incentive was the least important attribute (RI=0.77, 95% CI: 0.06–1.49). The RI for each attribute was statistically significant (all 95% CIs were greater than zero), yet considerable overlap of the 95% CIs reveal that several RI estimates were not significantly different from each other.

Fig. 2.

Relative importance of attributes on choices

Notes: Relative importance weights are conditional on the attributes included in the DCE and were estimated as the greatest difference in coefficients associated with each attribute.

Significant improvements in model fit when including interactions between the intervention and service delivery model attributes suggest that the effect of the interventions may vary, to some extent, according to the service delivery model. Allowing the effect of access to ART at the time of testing to vary by service delivery model approached statistical significance at a 5% level (p=0.06), and it is possible that a larger sample size would have rendered this interaction coefficient statistically significant.

3.3. Preference heterogeneity

The standard deviations estimated alongside the mean preference coefficients revealed significant heterogeneity in participant preferences for opting out of HIV testing (p<0.01), HIV testing at a community health campaign (p=0.05), obtaining an HIVST kit at a local pharmacy (p<0.01), and having access to ART at the time of HIV testing (p<0.01) (Table 3). The standard deviation for the interaction that allowed the effect of providing a testing incentive to vary by service delivery model was also statistically significant (p<0.01). While the mean preference coefficient for this interaction was near zero and insignificant, the relatively large and significant standard deviation reveals considerable variation in participant preferences for the interaction effect.

Model fit was not significantly improved by adding subgroup interactions to investigate whether preferences varied by participants’ self-reported HIV testing history, sexual behaviors in the 12 months prior to enrollment, or marital status (Appendix E). Model fit was significantly improved by incorporating subgroup interactions to investigate whether preferences varied according to participants’ perceived risk of having HIV (no or low risk versus moderate or high risk; LR test statistic=28.11, p<0.01), age (18–29 years versus ≥30 years; LR test statistic=21.66, p<0.01), and income (earns the daily median income or less versus above the daily median income; LR test statistic=19.46, p=0.01). We explored each of these subgroups further when simulating HIV testing uptake to assess how subgroup preference heterogeneity translates into differences in predicted testing behaviors.

3.4. Predicted uptake

Fig. 3 displays pre- and post-calibration predictions of HIV testing uptake by adult male residents when a single community-based testing service is made available. Pre-calibration predictions indicate much higher uptake rates than have been observed and confirm the importance of implementing the calibration procedure. Post-calibration predictions reveal uptake that ranged from a 0.15 to 0.91 share of the adult male population. Base case implementation of distribution of HIVST kits at local pharmacies (in which multi-disease testing, immediate ART access, and financial incentives are not provided) yielded the smallest predicted uptake share of 0.15 of adult men. Base case implementation of counselor-administered home-based testing (in which multi-disease testing, immediate ART access, and financial incentives are not provided) yielded a 0.33 predicted uptake share. Base case implementation of HIV testing at a community health campaign (in which multi-disease testing is provided yet immediate ART access and financial incentives are not) yielded a 0.64 predicted uptake share. Predicted uptake increased from the base case implementation scenarios by modifying the intervention attributes. Offering a US$ 0.85 incentive increased the predicted uptake share by 9–12 percentage points, depending on whether HIV testing was provided via pharmacy distribution of self-test kits or a multi-disease community health campaign, respectively. Providing multi-disease testing increased the predicted uptake share of counselor-administered home-based testing by 41 percentage points. Providing immediate access to ART increased the predicted uptake share by 24–43 percentage points, depending on whether HIV testing was provided via a multi-disease community health campaign or counselor-administered home-based testing, respectively. The greatest predicted uptake (a 0.91 share) was under HIV testing at a multi-disease community health campaign where immediate ART access and a US$ 0.85 incentive were provided (i.e. all interventions were implemented). Marginal increases in predicted uptake decreased as interventions were progressively added to a testing service, suggesting uptake is subject to decreasing returns to scale.

Fig. 3.

Predicted uptake when a single community-based HIV testing service is made available

MDT, multi-disease testing is available; ART, antiretroviral therapy is immediately available for HIV-positive persons

Predictions have been calibrated such that HIV testing uptake under base case implementation of a multi-disease community health campaign matches a 0.64 share of the adult male population. The calibration target is displayed in red.

Fig. 4 displays predicted HIV testing uptake by subgroups of men when a single community-based testing service is made available. Men who perceived their risk of having HIV to be moderate or high had higher predicted uptake of HIVST kits at local pharmacies compared to men who perceived no or low risk of having HIV. This finding held whether a US$ 0.85 incentive was not (a 0.31 vs. 0.14 uptake share) or was (a 0.49 vs. 0.24 uptake share) provided. Men who perceived their risk of having HIV to be moderate or high also had higher predicted uptake under base case implementation of counselor-administered home-based testing (a 0.69 vs. 0.29 uptake share) yet had lower predicted uptake under base case implementation of HIV testing at a multi-disease community health campaign (a 0.29 vs. 0.72 uptake share). One possibility could be that men who perceive greater risk of having HIV are drawn to the implicit privacy of the HIVST and home-based service delivery models. Young adult men aged 18–29 years had higher predicted uptake of HIVST kits at local pharmacies compared to men aged ≥30 years (a 0.69 vs. 0.14 uptake share without an incentive and a 0.46 vs. 0.30 uptake share with a US$ 0.85 incentive). Additionally, predicted uptake under base case implementation of counselor-administered home-based testing (a 0.78 vs. 0.37 uptake share) and under base case implementation of HIV testing at a multi-disease community health campaign (a 0.89 vs. 0.51 uptake share) were higher among young adult men. Predicted uptake under all testing alternatives was higher among men who earn the daily median income or less compared to those who earn above the daily median income. Particularly stark differences were observed in the predicted uptake of counselor-administered home-based testing. The predicted uptake by men who earn the median daily income or less was more than twice that of men who earn above the median income under base case implementation of counselor-administered home-based testing (a 0.55 vs. 0.23 uptake share) and increased to more than three times that of men who earn above the median daily income when multi-disease testing was provided (a 0.95 vs. 0.28 uptake share). These findings may suggest that accessibility to HIV testing and other health services is particularly important to men who earn the daily median income or less.

Fig. 4.

Predicted uptake by participant subgroups when a single community-based HIV testing service is made available

MDT, multi-disease testing is available; ART, antiretroviral therapy is immediately available for HIV-positive persons

Notes: Predictions have been calibrated such that HIV testing uptake under base case implementation of a multi-disease community health campaign matches a 0.64 share of the adult male population.

Table 4 displays the cumulative predicted uptake by adult male residents when two community-based HIV testing services are implemented concurrently. The cumulative predicted uptake ranged from a 0.50–0.96 share of the adult male population. Uptake increased as interventions were added to the base case service delivery models. Cumulative predicted uptake shares greater than 0.90 were achieved when at least one of the testing alternatives provided multi-disease testing or immediate ART access. The highest predicted uptake (0.96) was achieved when one of the services was a multi-disease community health campaign where immediate ART access and a US$ 0.85 incentive were provided (i.e. all interventions were implemented) and multiple interventions were also implemented via the other testing service. As with uptake under a single testing service, smaller marginal increases in predicted uptake were observed as interventions were progressively added to the two testing services.

Table 4.

Predicted uptake when two community-based HIV testing alternatives are made available concurrently

4. DISCUSSION

We predicted that providing community-based HIV testing services singly and in tandem can result in testing uptake by over 90% of adult male residents of rural Uganda. These findings are important given the need to expand HIV testing coverage among men in sub-Saharan Africa. Of the service delivery models investigated, HIV testing at a community health campaign and counselor-administered home-based testing demonstrated the greatest potential to generate broad uptake by men. Both service delivery models have been implemented and enhanced HIV testing uptake in Uganda [27,28]. It is therefore not surprising that men are willing to use these service delivery models. It is, however, unexpected that distribution of HIVST kits at local pharmacies generated the lowest predicted uptake given that high levels of acceptability with HIV self-testing have been reported elsewhere [31,32,62–64]. These findings could suggest that the particular pharmacy-based model of distribution holds limited appeal for the study population. Another possibility is that a lack of familiarity with HIVST kits as testing technologies diminished appreciation for the service delivery model.

We also found that attributes that represented interventions to promote HIV testing influenced choices. The largest increases in predicted uptake were observed when ART was immediately available for HIV-positive persons at the time of HIV testing. This finding is momentous given that many countries have recently adopted WHO’s “Treat All” guideline recommendation [65]. Recent studies have found that ART initiation on the same day of HIV diagnosis is feasible and improves ART uptake in facility-based settings [66,67]. Our results indicate that providing immediate ART access through community-based testing services is likely to also considerably increase HIV testing uptake. The provision of multi-disease testing also was a relatively important attribute to participants, and an inference that we could draw is that the provision of multi-disease testing contributes considerably to the high predicted uptake of HIV testing at community health campaigns. Implementers of the service delivery model have designed community health campaigns such that HIV testing is integrated with other testing services; we imposed an experimental design constraint to reflect this programmatic decision. Allowing the provision of multi-disease testing to vary across other service delivery models in the DCE allowed us to assess the relative importance of the multi-disease testing attribute, an achievement that has not heretofore been possible by observing HIV testing uptake at community health campaigns. The provision of a US$ 0.85 incentive also influenced choices. Recent experiments have demonstrated that incentives are effective at increasing HIV testing by adolescents and families in sub-Saharan Africa [68,69], and our results add evidence that financial incentives are likely to be an effective intervention to increase HIV testing by men. Although predicted uptake increases associated with offering an incentive were not as large as those associated with providing immediate ART access or multi-disease testing, the cost of the incentive was only US$ 0.85. A relatively small payment could be a cost-effective intervention to promote testing, and further research is needed to investigate the elasticity of testing behaviors in response to changes in the level of payment.

Considerable preference heterogeneity was detected by estimating a RPL model that allowed preferences to vary across participants, and exploratory subgroup analyses yielded compelling findings. For instance, we observed that the predicted uptake of HIVST kits at local pharmacies was higher among men who perceive a higher relative risk of having HIV. To the extent that perceived risk correlates significantly with actual risk of having HIV, distribution of HIVST kits at local pharmacies could yield more new HIV diagnoses, despite lower uptake among the adult male population at large. Additionally, we observed that men who earn the daily median income or less had higher predicted uptake of all community-based testing services presented in the DCE than men who earn above the daily median income. This finding may be important for considering how to ensure broad HIV testing access given that travel distance and poverty have been cited as barriers to facility-based testing [70–72].

We acknowledge several study limitations. First, the generalizability of our results is limited to adult male residents of rural Uganda (i.e. 75% of the adult male population). Additionally, comparison of our sample characteristics to nationally representative survey data revealed that young adult men (aged 18–29 years) were underrepresented in our analyses. Our exploratory subgroup analyses suggested that the preferences of young adult men may differ somewhat from those of older men; in particular, distribution of HIVST kits at local pharmacies appeared to appeal more strongly to young adult men, and our results are likely skewed toward the preferences of older men. Second, we elicited preferences for HIV testing attributes that evidence suggests hold potential to increase male testing and that are highly relevant from a policy perspective, yet we did not conduct formative research prior to the DCE to establish attributes that are important to the study population. Additional attributes may influence men’s decisions to test for HIV. We assumed that attributes that were not defined in the DCE or standardized by the introductory script were constant across testing alternatives, and our calibration procedure adjusted for unobserved factors that influenced choices to opt-out versus opt-in to testing. We acknowledge that our results could have been affected if participants associated unspecified influential attributes with attributes that we included in the DCE. Third, our experimental design was tailored to our study objectives, took into account constraints of our study environment, and leveraged prior information about preferences yet our manual design construction was unconventional. Calculation of J-index correlation coefficients indicate two high correlations in the differences in the levels of the attribute interaction terms and attribute variables. We could have chosen to use of final model specification that did not include attribute interactions, yet we found that the attribute interactions considerably improved model fit and therefore interrogated whether the correlation was severe enough that it produced signs of multicollinearity by estimating VIFs. The VIFs indicated no signs of multicollinearity, yet we nonetheless acknowledge that it is possible that another design could have eliminated such concern. Further, our design was not powered to detect the statistical significance of interaction terms at conventional thresholds. We based our inclusion of attribute interactions in our final model specification on improvements in model fit. Similarly, we explored uptake by participant subgroups for whom subgroup interactions improved model fit. We thus were able to use our data effectively to estimate preferences and conduct exploratory participant subgroups, yet a larger sample size or more choice set questions would have been beneficial to assess the statistical significance of the interactions. Fourth, we investigated preferences among a population-based sample to respond to the challenge of low population-level uptake of HIV testing by men, yet another approach to ensure that 90% of men who are living with HIV know their HIV status is to target subgroups of men who are at high risk of acquiring HIV. As progress toward the 90–90-90 targets continues, ensuring not only broad population uptake but uptake by subgroups at risk of acquiring HIV will likely become increasingly important. Although we discovered improvements in model fit when investigating whether preferences varied by participants’ self-perceived risk of having HIV, we were not able to fully clarify the preferences of high-risk subgroups in this study.

Despite limitations, this study provides timely, useful guidance for decision makers who seek to expand HIV testing coverage among men in sub-Saharan Africa. We also contribute several methodological advances to the stated preference literature that are especially beneficial for investigators who seek to conduct applied analyses to inform decisions regarding the design and delivery of health services. First, when evaluating the correlation properties our experimental design, we made the distinction to evaluate correlation with respect to differences in attribute levels in accordance with how discrete choice models are derived. Evaluation of the correlation properties of experimental designs has been guided by metrics that were established for linear model estimation, and we took a step to adapt one such metric for the nonlinear model estimation of stated preferences. Second, we discuss how we employed a hybrid coding scheme that ensured that attributes variables and interactions were not confounded with the constant for the opt-out alternative. We are not the first authors to employ a hybrid coding scheme, yet we advance the literature by describing our methods and the particular interpretation the hybrid coding scheme achieved. Third, we describe how we conducted subgroup analyses using interaction terms. Such an approach has not been widely used in stated preference studies despite great usefulness to overcome issues related to scale differences that arise when estimating separate regressions [60]. Fourth, we implemented a calibration procedure using participant-specific preference estimates to ensure that our uptake predictions were consistent with reference values of observed uptake that have been reported in peer-reviewed literature. Calibration is not routinely performed in the stated preference literature, yet we contend it is a critical step to advance the use of DCEs to predict service uptake.

Finally, our work motivates further investigation. Further research is needed to investigate the feasibility of providing ART through community-based testing services and to identify effective strategies to link HIV-positive persons to long-term treatment and care. We found considerable preference heterogeneity, and research to determine how to optimize HIV testing uptake by priority subgroups (both male and female) would be useful. Additionally, we here predict the impact that implementing promising certain service delivery models and interventions is likely to have on testing uptake by men, yet additional research is needed to evaluate the cost-effectiveness and long-term health outcomes of community-based strategies to promote testing. Collectively, these actions will strengthen evidence regarding how community-based services can be optimized to promote testing and support continued progress toward the 90–90-90 targets.

KEY POINTS FOR DECISION MAKERS.

• We predict that providing community-based HIV testing services singly and in tandem can result in testing by over 90% of adult male residents, yet uptake strongly depends on attributes of testing services.

• Ensuring immediate access to antiretroviral treatment at the time of HIV testing is likely to have a large, positive impact on uptake of community-based testing.

• Considerable preference heterogeneity exists and predicted uptake differed by participant subgroups; further research is needed to clarify how to optimize testing to promote uptake by priority subgroups.