Different attributes and the levels of goods and services determine the demand for them. This study applied the DCE method and guidelines that Lancsar and Louviere (2008) recommended, based on the responses from 275 antenatal care (ANC)-attending pregnant women. The methodological steps were conceptualising the choice process, defining attributes and levels, experimental design construction, sample size, design blocking and the econometric model.
2.1 Conceptualising the choice process
Qualitative semi-structured interviews were conducted with health sector decision-makers at different levels to develop fair choice sets and conceptualise the overall process. Decision-makers/healthcare managers were purposively selected as key informants. The following inclusion criteria were set: 1) hold a responsible position in making HIV/AIDS/PMTCT-related healthcare decisions (i.e., general director, vice-director); 2) lead a PMTCT service planning, financing, monitoring and evaluation department; or 3) be involved in day-to-day PMTCT service delivery to pregnant women. Decision-makers at central, regional, zonal and district levels were interviewed about their views of DCE choices and overall economic evaluation, practice and barriers.
A total of 48 key informants were interviewed from the Federal Ministry of Health (FMOH) down to the health facility level. To reach saturation, an additional eleven respondents working at either the Regional Health Bureau, zonal health departments or district health offices were enrolled. All the interviewees held the organisational head or vice-head position or were management team members. In particular, they explained their views on the DCE attributes and levels. Table 1 sets out a detailed profile of the study participants.
Table 1
Decision-makers/healthcare managers from the central to the healthcare provider level for the DCE and economic evaluation preliminary analyses
Key informants (for general health service) | Number (%) of respondents |
At the regional level (RHBs) | 6 (13%) |
At the zonal and district level | 5 (10%) |
At the health facility level | |
Hospital CEO | 4 (8%) |
Health centre heads | 8 (17%) |
Healthcare provider | 12 (25%) |
Key informants (for PMTCT program) | |
At the central level (FMOH, EPHI, FHAPCO) | 2 (4%) |
At the regional level (RHBs) | 6 (13%) |
At the zonal and district level | 5 (10%) |
Abbreviations: RHBs: Regional Health Bureaus; FMOH: Federal Ministry of Health; FHAPCO: Federal HIV/AIDS Prevention and Control Office and CEO: Chief Executive Officer |
[Insert Table 1]
To highlight the program-specific context, two PMTCT national program coordinators (at the central level), six PMTCT regional level program coordinators (in the four regions and two city administrations) and five zonal/district PMTCT program coordinators were included in the qualitative interviews. Facility heads and health workers from twelve health facilities were also interviewed. At the lower health service tier, fourteen health facility administrators/chief executive officers and facility heads were enrolled for the qualitative interview sessions. Furthermore, twelve health experts working in ANC, maternal and child health, and labour and delivery services were interviewed to understand the critical choice attributes and levels for the preference analysis. The qualitative findings on economic evaluation across health sector decision-making, and from the PMTCT program perspective, were separately reported (Zegeye EA et al., 2017).
Four possible attributes, namely, waiting time, service fees/money, location and service integration, were presented using the semi-structured interview guide. The guide asked the respondents if any relevant attributes and levels were missing for the DCE choice set construction. The guide contained questions that sought to define the role of the key informant, provide an economic evaluation overview, explain the current challenges associated with low PMTCT coverage, and identify the possible attributes/factors associated with PMTCT service utilisation and preferences among pregnant women. Based on the findings from the qualitative interview sessions, the relevant DCE attributes and levels are discussed in the following sections.
2.2 Attribute selection and defining levels
Based on the key informants interviewed, the PMTCT service attributes and levels were defined. The four attributes defined as waiting time, cost/service fees, location, and service integration, were presented to verify/cross-validate whether these service factors symbolised the PMTCT service set-up/context and would help to elicit pregnant women’s preferences. The key informants agreed that waiting time, cost/money and service integration should be considered in the DCE; however, the location attribute and its levels were revised based on the feedback from the interview. Moreover, the key informants strongly recommended the addition of two new attributes: pretest counselling and post-test counselling.
After thorough interview sessions and qualitative analysis, three attributes with four levels and three with two levels were developed for the DCE experimental designs. Based on these attributes and levels, hypothetical DCE choice sets were constructed, assuming that the stated choices would elicit pregnant women’s preference towards different PMTCT service package alternatives. The detailed attributes and levels (including the baseline comparators) are summarised in Table 2.
Table 2
Selected attributes and levels for the DCE experimental designs
DCE attributes | Definition of Levels |
1 (Baseline) | 2 | 3 | 4 |
Service integration | At ANC/ Labor & delivery | ART/ family planning | | |
Pretest counselling | Be counselled individually | Receive couple counselling | | |
Disclosure counselling | One-on-one post-test counselling | With a partner post-test counselling | | |
Waiting time | 30 minutes | 45 minutes | 1 hour | 2 hours |
Cost/Service fees | Free | You pay 25 ETB | You pay 50 ETB | We pay 100 ETB |
PMTCT site location | At health post | At health centre | At district hospital | At specialised hospital |
[Insert Table 2]
The final six attributes developed were service integration, pretest counselling, disclosure counselling, waiting time, cost/service fees and PMTCT site location. Service integration refers to the placement/integration of the PMTCT service within the health facility service delivery setting/compound. It could be integrated with ANC, the labour and delivery service (first level), or Antiretrovial treatment (ART) or family planning services (second level). The second attribute, pretest counselling, involves the provision of relevant HIV-related education and information for pregnant women or their partners before proceeding to the HIV testing service. This attribute has two levels: individual pretest counselling and couple counselling. The third attribute, disclosure counselling, deals with the provision of relevant, confidential and private post-test HIV results, counselling and advice for the patient (or a couple). This attribute consists of two alternatives: one-on-one post-test counselling or with a partner. The fourth attribute, waiting time to receive PMTCT services, was identified as a relevant attribute, and its levels consist of waiting 30 minutes, 45 minutes, 1 hour and 2 hours. The fifth and sixth identified attributes were cost/service fees and PMTCT service location. Cost/service fee refers to the amount charged for PMTCT services. The defined levels for this attribute were free of charge, 25 ETB (USD 1.27), 50 ETB (USD 2.54) or receiving 100 ETB (USD 5.08) as compensation/to incentivise and improve service. The PMTCT service location attribute provides service across the different health tiers (i.e., at the health post, the health centre, the district hospital or specialised hospital). Each tier has its own defined catchment population[1].
2.3 Experimental design construction
The experimental design step is the foundation for the choice set construction and enables a proper estimation matrix. The design considered different combinations of attributes and defined levels, which ultimately led to the production of choice sets. There are two dominant experimental designs: full factorial and fractional. The full factorial design applies the LM formula, where L refers to the number of levels and M refers to attributes (Lancsar and Louviere, 2008). Accordingly, in this study, three attributes with four levels and three with two levels were inputted into the formula. The total number of full factorial choice sets would be 512 = 43 X 23. However, due to the feasibility challenges associated with the data collection, the fractional experimental design was selected as an appropriate approach where samples were used from the full list of choice sets (Louviere et al., 2000).
Although the fractional design is believed to reduce efficiency, it has recently become more popular in the literature by applying statistical efficiency measures (de Bekker-Grob et al., 2012). Small fractional factorial designs (orthogonal main effects plan/OMEDs) were developed to show the main effects (the relationship between the outcome variable and the independent variable) and possible interaction effects (the effect of one attribute on the other) of variables (Lancsar and Louviere, 2008). Within this framework, this study employed OMEDs designs, which were generated using the SPSS software. The overall design formulation followed the factorial design approach that Street et al. (2005) reported. The theoretical analysis of alternative factorial design strategy 5 (applying OMEDs and search algorithm) is the recommended design that will result in an optimal, or near-optimal, combination of alternatives, which will be able to detect the main interaction of the variables in the model.
This design was also consistent with the four criteria of an efficient design: orthogonality, minimum overlap, level balance and utility balance (Huber and Zwerina, 1996). An orthogonality design comprising 64 scenarios was generated using the SPSS (Louviere et al., 2000). The software generated choice sets that ensured each selected attribute was independent of the others and hence a statistically guaranteed level balance (equal representation of levels with equal frequency) in the choice sets. The property of minimum overlap was also pursued, following the alternative option construction. Once the software generated Option A alternatives, the second alternative (Option B) was developed by cyclically adding one additional level. This choice set computing technique finally led to zero overlaps of levels (minimising the probability of level repetition in the choice sets). Finally, the utility balance was assumed to be equal across the alternatives (Options A and B). This is because all the respondents (pregnant women attending ANC services) have equal PMTCT service packages (despite different attributes and levels), which would be expected to yield equal utility. Applying these procedures, the DCE design fulfilled the four efficient design criteria.
Finally, in accordance with the DCE procedures, after the possible list of 64 choice sets (and alternatives: Options A and B) was determined, the next step was blocking the design. This helped to ensure a fair level of distribution of questions per respondent (de Bekker-Grob et al., 2012). Four blocks (A, B, C and D) were formulated and mainstreamed in the OMEDs’ orthogonality choice set design in Annexure 8 without affecting statistical design efficiency. In each surveyed health facility, a minimum of twenty pregnant women, five/six respondents per block, were interviewed. Following this procedure and four block designs, for the total of 275 pregnant women attending ANC/PMTCT services, 5,366 observations were considered for the DCE analysis.
2.4 Econometric model
In the context of DCE, the literature discusses different econometrics modelling (de Bekker-Grob et al., 2012; Lancsar and Louviere, 2008). McFadden’s multinomial logit (MNL) (1974) was the foundation and prominent model fitting the discrete preference choice. The other models originated from MNL’s three important, basic assumptions of individual behavioural choices: independence irrelevance alternative (IIA), identical distribution of the error terms across observations (IID) and homogenous preference across DCE respondents (de Bekker-Grob et al., 2012). The other recent econometric models were developed by relaxing the above assumptions. These include the nested logit model, multinomial probit models, heteroscedastic logit, mixed logit, conditional (fixed effect) logit and the random effect logit model. Health sector studies have started applying alternative models to analyse the relevant attributes/levels determining choices (Lancsar and Louviere, 2008). Recently, flexible econometric models (of the three MNL assumptions) have increasingly been applied in the healthcare sector.
Based on maximum likelihood estimation (MLE) approaches, the conditional (fixed effects) and random effects logit have commonly been employed in the literature. Derived from the consumer choice theory, the MLE assumes individual choice based on a service/good’s attributes and level, providing a higher utility level. Hence, the MLE estimates the probability of an individual choosing preferable attributes (and levels) compared to alternative choice sets. The two recent models (conditional/fixed effects and random effects logit) differ from the above two assumptions of IIA and IID (Han, 2004). These models relax these assumptions and are more appropriate (descriptive analyses) for the data sets. Accordingly, consistent with the literature, conditional (fixed effects) and random effect logit models were employed in this study to estimate the likelihood and odds ratio of different alternative attributes and levels (6 attributes and 2(4) levels) with a dependent variable (i.e., choices). These models were tested using the Hausman test (Hausman and McFadden, 1984).
The dependent variable was the choice (coded 1 if the pregnant woman chose the alternative with the given attributes (or levels) as compared to the baseline comparator, that is, 0). Applying the logit model estimated the probability that the respondents (pregnant women attending ANC/PMTCT services) would choose the given attribute level in Option A compared to the different choices set in Option B. As noted in the previous section, the choice sets were fairly distributed based on the orthogonal design. The explanatory (independent) variables were the different hypothetical alternative attributes and levels of the PMTCT service packages (service integration (level 1, level 2); pretest counselling (level 1, level 2); disclosure counselling (level 1, level 2); waiting time (level 1, level 2, level 3, level 4); cost/service fees (level 1, level 2, level 3, level 4:) and location (level 1, level 2, level 3, level 4)). The details are set out in Table 2.