Designing a package of sexual and reproductive health and HIV outreach services to meet the heterogeneous preferences of young people in Malawi: results from a discrete choice experiment

Uptake of family planning (FP) and HIV services among sexually active youth in sub-Saharan Africa remains low. High levels of unmet need for FP is associated with increased adolescent pregnancy leading to both increased morbidity and mortality [1-4]. There have been repeated calls to increase availability, accessibility and quality of sexual and reproductive health (SRH) services, including FP, for young people in low-income settings [4-7]. However, the evidence directly linking increased availability of youth-friendly services and increased uptake of FP services among youth is mixed [8,9] suggesting that there may be additional barriers to access or heterogeneity in youth preferences which may not have been addressed in the design and delivery of services for young people.

In addition to the development of youth-friendly services, recent efforts to improve access to FP and HIV services in low- and middle-income countries have focused on service integration. At the operational level, the term integration may include efforts to join FP and HIV services through the coordination of financing and policy. At the service delivery and facility levels, integration may refer to methods of coordinating service delivery to allow individuals to access multiple services in a single consultation, either by the same provider or on the same site through a system of internal referral [10]. Both types of integration have the potential to improve access to FP services, reduce stigma associated with accessing HIV services and increase patient satisfaction. However, available evidence to date shows that the extent to which this potential is realised through service delivery and facility level integration is mixed; missed opportunities to provide multiple services in a single visit persist [11-16], and the complex issues that surround stigma and patient satisfaction are not adequately addressed in many integrated service delivery configurations [15,17,18]. Resource constraints, high patient volumes and inadequate provider training contribute to difficulties in effectively integrating services, but despite these challenges there is consensus among patients and policy makers alike regarding the need to better integrate FP and HIV service delivery and to understand how to design services that are responsive to client needs [14,19].

In order to develop a patient-centred approach to the design of integrated FP and HIV services for young people it is important to identify which aspects of service delivery youth value, to understand the relative importance of service characteristics and to understand how preferences are likely to vary among different groups. To date, no empirical work has assessed patient preferences for integrated FP and HIV services or examined the impact of the service package configuration on young people’s preferences. We used a discrete choice experiment (DCE) to assess young people’s stated preferences for the design and delivery of integrated FP and HIV outreach services in rural Malawi. The research was conducted in partnership with the Family Planning Association of Malawi (FPAM), a local non-governmental organisation providing both facility based and outreach services in four districts in Malawi. In line with the government focus on improving access to SRH and HIV services to young people, FPAM is working to expand outreach service provision including the delivery of FP and HIV counselling and testing (HCT) in the four districts it is currently operating in. The present study was designed to inform the design and delivery of outreach services though this expansion.

The aim of this research was to identify the relative importance of service characteristics; to identify additional youth-friendly service components that may be attractive for young people; to understand preferences for integrated FP and HIV service delivery; to explore heterogeneity in youth preferences; and to estimate willingness-to-pay (WTP) for service attributes. It was hypothesised: (i) that male and female respondents would have different preferences for service attributes, (ii) that younger respondents would have less well formed preferences and that this would be revealed through heterogeneity in preferences for service attributes^a and (iii) that individuals who have experience in using FP or are currently in a relationship would have different preferences for family planning services.

DCEs are a multi-attribute survey method used to indirectly elicit consumer preferences by presenting respondents with hypothetical situations which may be seen as reasonable substitutes and asking them to choose between alternatives [20]. Using this approach, a health service may be described by a number of characteristics or attributes. Choice profiles are then created with varying levels of each attribute and paired or grouped to create choice sets. Respondents are asked to make repeated choices over a number of different choice sets in order to observe the tradeoffs that they make as the levels of each attribute varies. The value that respondents place on each of the good’s attributes is inferred by calculating the marginal rate of substitution between attributes [21]. Where price is included as an attribute, WTP for a change in attribute levels may be estimated as the negative ratio of the attribute and price parameters.

In addition to allowing for the exploration of preferences for services which are currently not available, DCEs may include attribute levels that are outside the range of those that patients normally encounter [22]. This can be helpful for understanding the impact of changes to current services and may be helpful in identifying strategies to make services more attractive to clients. Such analyses can inform the design and delivery of services or interventions and may be helpful in identifying strategies to increase the acceptability or uptake of existing or new services.

DCEs are commonly used in health economics to elicit patient preferences for a variety of services and health care delivery models [23-26]. They have also been used to explore preferences for and predict uptake of existing and new products [27-29]. The majority of published DCEs in health economics have been conducted in high income settings and with adult populations [30]. Few examples of DCEs done with young people have been published and these were conducted in high income settings [27,31,32]. As a result, little is known about the use of DCEs among young people in low- and middle-income settings. This study provides the first example of a DCE conducted among young people in a low-income setting.

The results of the analyses presented here suggest that when it comes to FP services, respondents value confidentiality more than any other service attribute. This result is not surprising and is consistent with the qualitative work done in developing this DCE and the literature available on youth preferences for SRH services [45,46]. In this setting confidentiality was described as a service where clients would feel confident that no one would overhear their consultation and the service provider would not divulge the details of the visit to anyone else. This definition covers concepts of privacy and secrecy but does not cover anonymity which is an element of confidentiality that young people may value [46], but which may be hard to ensure in an outreach setting.

Perhaps a more unexpected result is the importance of the HIV service package that is available in the context of FP service delivery. Respondents clearly value the availability of both HCT and ART services more than the presence of HCT alone. Given the low proportion of youth in this setting who are likely to be HIV positive and know their status, this suggests that youth valued these additional services regardless of whether they would actually use them. This suggests the value that youth place on this attribute may be more akin to an ‘option’ value [47] meaning that respondents place value on having the option to use this service, even though they may not use the service themselves. The positive value that youth place on the addition of ART to the HIV service package suggests that the availability of this service would not stigmatise the other services provided at the same time which can be a concern when considering SRH and HIV service integration [48]. This is supported by the findings from the qualitative work in which young people expressed concern that ART services should be readily available for those receiving a positive HIV test result.

The importance of the scale parameter clearly points to significant scale heterogeneity in responses. This type of heterogeneity cannot be explained by observable respondent characteristics and it is often unclear what the source is. Sivey et al. [49] have suggested that respondents who have difficulty in choosing, who had a hard time understanding the choice tasks or who have very low marginal utilities for some attributes may exhibit choice behaviour that appears to be nearly random which may result in a significant amount of heterogeneity in the scale parameter. Fiebig et al. [29] found this kind of behaviour to be more common in choice experiments with complex designs or those involving medical decision making. However, at present it is not possible to determine precisely which of these hypothesised sources of heterogeneity is driving this.

Given the unlabelled design of the DCE which presents two generic hypothetical situations and the framing of the choice tasks which asks respondents to imagine that they have already decided to access services, the results of the probability analysis and the proportion of respondents expected to choose a given service package are unlikely to correspond directly to increases in service utilisation. Simulated changes in probability also reflect in part the market simulated by the experimental design, which may not correspond directly to the actual market. As a result, these estimates are better understood as a measure of the acceptability of outreach services and may be used to understand which scenarios are most likely to be attractive or acceptable to young people.

Despite the significant amount of heterogeneity revealed in the model estimates, the results of the probability analysis consistently show that respondents prefer a service package that offers confidential services, both HCT and ART and sports for youth, with 29-32% of respondents expected to choose a service package including these three elements. Where service providers are unable to offer sports and HIV services, simulations indicate that 20% more respondents are expected to choose a confidential service compared to a service where clients may have concerns about confidentiality. This may offer some important insight to service providers aiming to improve the acceptability and youth-friendliness of their FP service provision with limited resources; where it is not possible to add additional service components, acceptability may still be improved by prioritising the elements of FP services that are most important to young people. This may be particularly important in areas where researchers have found that increasing the availability of youth-friendly SRH services is not sufficient to ensure that uptake increases among young people [8,9].

Estimates of WTP indicate that young people would be willing to pay for some service attributes. However, the results of the simulation analysis illustrate the impact of increasing price on young people’s preferences through the decreasing proportion of young people expected to choose a scenario decreases as price increases, with 12% fewer respondents expected to choose a service costing 500MK compared to a free service. While WTP is an informative measure and is helpful in terms of quantifying the relative importance of attributes, there are several reasons that it should be interpreted with caution in this setting. Firstly, within the survey sample the majority of respondents reported a monthly household income of USD $4 – 40, meaning that an estimated WTP of USD $1.76 for a confidential service could potentially be a large proportion of household income for many respondents. Secondly, while many young people in the sampled age range of 15–24 years may already be head of household and the majority of respondents indicated that they were engaged in some form of paid employment in the 12 months preceding the survey, many respondents in this age group would have been living with parents and many were not engaged in any form of employment. These respondents are unlikely to have direct access to their own financial resources and only indirect access to household resources. Many young people may therefore need to ask for financial resources to pay for services or for transportation to access services. Thirdly, young people in general, and particularly young women, may have limited capacity to make independent decisions about accessing FP services. Some young people may access services without their parents’ or partners’ knowledge, but making arrangements to attend services in secret may present a barrier to access for young people. Similarly, married women in this age category may be unable to make independent choices about contraceptive use, particularly in a cultural setting where child-bearing is highly valued and infertility stigmatised [28,29]. Finally, of 269 respondents who had previously accessed FP at either at a health facility or using an outreach service, over 80% reported receiving FP free of charge, indicating that very few respondents had experience in paying for FP services. As a result, drawing policy conclusions or pricing strategies based on estimates of WTP may prove problematic.

Overall, similarities in age and gender structure between the study sample and the general rural population in Malawi may mean that some results could be generalised to other rural settings in Malawi, particularly where analyses have revealed heterogeneity in preferences according to these characteristics. The relatively high reported rates of FP use among 15–19 year olds in the study sample may be a result of over reporting or it may mean that respondents in the study sample have better access to FP and face fewer barriers to access than young people in other areas of the country. If actual FP use is higher in the study population than the general population in rural Malawi then respondents may have more complete preferences and estimates of the proportion of respondents expected to choose specified service packages may be higher in this population than if similar work was completed with other groups. However, the results of this analysis did not indicate that previous FP use had a significant impact on preferences; indeed this was not included as an interaction term in the final model specifically because it was not a significant source of heterogeneity in any of the exploratory model specifications.

Respondents completed a complex choice experiment and appeared to trade between outreach service attributes. There is considerable variation in young people’s preferences for FP service characteristics but targeting a few key areas of service design and delivery may result in similar improvements in service acceptability across all the sub-groups included in the analysis.

Further work could be done to understand the sources of heterogeneity and to determine how this may impact optimal service design and delivery strategies. This type of analysis may be of particular interest to policy makers interested in targeting a particular sub-group, such as girls or adolescents. This work could also be expanded to look at young people’s choice of SRH service provider in order to better understand preferences for outreach services compared to clinic-based services.

^aRespondents who do not have well-formed preferences may learn about their own preferences through the process of completing the choice tasks, or may construct preferences around the attributes presented in the choice tasks over time. These respondents are likely to be less consistent in their responses over time and this variation can show up as large or significant standard deviations around mean parameter estimates [50].

^bIn contrast to orthogonal experimental designs which aim to minimise correlation between attribute levels, efficient designs aim to produce data with as low as possible standard errors. Such designs incorporate prior parameter estimates to determine the asymptotic variance covariance (AVC) [51]. A d-efficient design is one where the determinant of the AVC is minimised. This is the most commonly used approach and the method used in both the pilot and final designs.

^cThis approach was used in order to create a design that could be used for either a forced choice or unforced choice format and analysed using an MNL or RPL approach. This involved specifying four utility functions in Ngene and using the ;model command to calculate the AVC matrix for each specification. Ngene then minimises a weighted average measure of efficiency across each model [51,52]. Maintaining this flexibility at the outset meant that both choice formats could be included in the questionnaire and participants who opted out could then be asked a follow-up forced choice question. In the event that many participants opted out this approach would allow for a complete forced choice dataset that could still be analysed. Since very few participants opted out in the final survey the unforced choice data are used for the present analysis.

^dGuidance in the literature suggests that the first step in determining the sample size for a DCE is to determine the expected proportion of respondents choosing the least popular alternative. Using the unlabelled DCE with two alternatives (Service A and Service B) and an opt-out as a guide, it was estimated that roughly half of respondents would choose neither of the alternatives proposed (meaning they would choose the opt-out alternative) and the rest of the respondents will be equally split between the two. The second consideration is the number of choice sets to be completed by each respondent. Prior to the DCE development, it was anticipated that a maximum of 16 choice sets would be presented to each respondent in order to avoid overburdening respondents with too many choices. Finally, sub-group analyses were considered. In order to identify variation in preferences according to age and gender 4 sub-groups were identified; young women aged 15–19 years, young women aged 20–24 years, young men aged 15–19 years, young men aged 20–24. Based on these parameters, it was estimated that a minimum of 150 respondents for each of the sub-groups identified, or a total number of 600 (150 x 4) would be required. If each of the 600 respondents completed all 16 choice sets then 9600 observations (600 x 16) would be included in the data set. If 50% of the choices were ‘opt-outs’ then a remaining 4800 observations would still be available to analyse tradeoffs between the various attributes included in the service profiles. The 2010 DHS obtained a response rate of 98% and noted that response rates tended to be higher in rural areas [44]. Assuming that similar response rates could be achieved in the study sites, and assuming that each respondent would be able to complete the survey it was estimated that in order to obtain 600 completed surveys, approximately 620 individuals would need to participate in the survey.

^eEffects coded variables use −1 as the base category instead of 0. For a variable with only two levels, this would be coded as −1 when the characteristic is not present and 1 when the characteristic is present. For three or more levels −1 remains the base category and the variable is coded as 1 when the characteristic present and 0 otherwise. This coding format allows effects to be separated from the overall mean effects measured in the utility function [22,53].

^fFiebig et al., [39] found that AIC correctly identified GMXL with correlated parameters as the true underlying model structure in 100% of model simulations.

^gThe Independence of Irrelevant Alternatives (IIA) assumption is not imposed in the RPL or GMXL framework as it is in an MNL model meaning that changes in probabilities associated with a change in more than one attribute are not a linear combination of changes estimated for individual attributes.