A model based cost-utility analysis was undertaken to evaluate the costs and benefits associated with DAART in comparison with self-administered ART using a life time Markov model with half cycle correction [16, 17]. The study patient population was HIV-positive infected adults in sub-Saharan Africa, with a CD4 count > 500mm³, classified as being at high risk of non-adherence to ART. Patients were assumed to be 20 years old (the impact of this assumption was examined through sensitivity analysis) and the cycle length was one year. The perspective of government health care payers in a sub-Saharan Africa setting was taken, with only direct costs incurred by the health care provider being included in the model. The model structure consists of five HIV health states, based on the clinical categories of patient CD4 cell counts (Fig 1): State I, is HIV infected adults with a CD4 count of greater than 500 cells/mm³, without any symptoms; State II is CD4 counts of greater than 350 cells/mm³ but less than or equal to 500mm³, and is categorised as mild asymptomatic; State III is CD4 counts greater than 200 cells/mm³, but less than or equal to 350 cells/mm³, categorised as symptomatic; State IV is CD4 counts less or equal to 200 cells/mm³, categorised as severe symptomatic /AIDS; and State V is death. Patients can remain in the same state, or transition to any worse state or better. This approach is supported by evidence which has demonstrated the possibility that patients with AIDS can improve and move to a better HIV state, including the HIV asymptomatic state [18]. Half cycle correction was applied to costs and benefits [16, 17]. Costs and utilities were discounted at 3% [19, 20]. We used the WHO-CHOICE threshold of US$5,086 as the basis for assessments of cost-effectiveness [19, 20].

Data used in the model were derived from the published literature and are summarised in Table 1 and described below.

All the patients were assumed to start in ‘State I’ and thus have a CD4 count greater than 500 [21]. For HIV/AIDS disease progression under the self-administered ART arm, data were used from patients that were followed-up between 2005 and 2009 from a referral hospital in Ethiopia (Table 1) [21]. The transition probabilities between health states in the directly administered arm were based on an estimate of the treatment effect of DAART compared with self-administered ART (relative risk) derived from a recent meta-analysis of 14 controlled trials (both randomized and non-randomized) [5]. The effect of this relative risk (RR) was calculated for each transition probability as follows [22]:

Age-specific all-cause (not due to HIV/AIDS) mortality was applied to each state in the model (Table 2), and was calculated using the age-specific mortality rate from sub-Saharan Africa. This was applied in addition to the death rate from HIV/AIDS [23].

All the utility values for each HIV/AIDS health were derived from a survey conducted in South Africa using validated questionnaires, with HIV patients at different stages of HIV disease progression (Table 1) [24–27].

The costs in this analysis included treating opportunistic infections, and all the costs of inpatient and outpatient health care utilization (Table 1) [24, 28–31]. The annual inpatient days and outpatient visits were calculated based on annual mean inpatient days and outpatient visits reported in the literature (Table 1)[32]. All the costs are converted to USA dollars, and inflated to 2014 prices [33].

A series of sensitivity analyses were carried out to examine the impact on the model results of the uncertainty in the model input parameters. Structural uncertainty was assessed by assuming that no reverse progression of HIV/AIDS patients was possible (i.e. from a bad state to a better state). One way sensitivity analysis was performed, by increasing and decreasing all the input parameters by 25%, in order to examine which parameters have the greatest influence on the model results. Probabilistic sensitivity analysis (PSA) was performed with the model parameters varied simultaneously according to pre-specified distributions. Dirichlet distributions were used for all transitions with more than 2 possible outcomes. Beta distributions were used for all transitional probabilities with only two options, utilities and HIV state transition probabilities. Gamma distributions were assigned to all costs and a lognormal distribution was assigned to the effectiveness relative risk for DAART. 10,000 Monte Carlo simulations were implemented with the results shown using a cost-effectiveness acceptability curve which provides a measure of the likelihood that a decision to apply a given intervention is cost effective across a range of ‘willingness-to-pay’ thresholds, where the willingness-to-pay represents the maximum amount decision makers are willing to pay for a unit of QALY gain.

The (individual patient) expected value of perfect information (EVPI) was calculated as the difference between the expected net health benefit (NHB) given full information and the expected NHB given current information. The NHB measure is the increase in effectiveness multiplied by the amount that the decision-maker is willing to pay per unit of increased effectiveness (i.e. the cost-effectiveness threshold), minus the increase in cost. The population EVPI was estimated by multiplying per patient EVPI by the effective population, i.e. the annual population of at high-risk of non-adhering to ART discounted over the lifetime of the treatment (assumed to be 50 years). Among people living with HIV in sub-Saharan Africa (≈24.4 million [1]) about 17.4% [34] were at high-risk of non-adhering to medication, and only 37% of people living HIV in sub-Saharan were on ART in 2013 [1]. The effective population in sub-Saharan Africa was therefore estimated at 1,600,000. The EVPI represents the maximum amount that a decision-maker should be willing to pay for further information to guide the adoption decision in the future, where additional research should be considered if the EVPI exceeds the cost of research. The expected value of partial perfect information (EVPPI) was also calculated in order to provide more focus for further research, by identifying the groups of inputs where it would be valuable to have more accurate estimates. Specifically, EVPPI estimates were generated for the following sets of parameters: effectiveness, cost, utility, and probabilities. The model used for the EVPPI was assumed to be linear for ease of computation.

The base case results of the model are summarized in Table 3. The expected cost of self-administered ART and DAART were $5,200 and $15,500 respectively and the QALYs gained were 8.52 and 9.75 for self-administered ART and DAART respectively. The incremental cost per QALY gained for DAART versus self-administered ART was $8,400 per QALY, this suggests that DAART is not cost-effective based on the WHO-CHOICE threshold of US$5,086. The results of the simplified model assuming no reverse disease progression yielded an ICER of $9,700 per QALY.

The results of the one-way sensitivity analysis are shown in Fig 2. The estimated ICER was most sensitive to variations in the cost and effectiveness of DAART. When the annual cost of DAART was varied from US$723 (best case scenario) to US$1,205 (worst case scenario) per annum, the estimated ICER ranged from US$6,440per QALY gained to US$10,921. Similarly, when the effectiveness of DAART was varied using the highest estimate from the literatre (RR = 1.49, best case scenario) and the lowest (RR = 1.12, worst case scenario), the estimated ICER was US$6,065.01 per QALY and US$17,987 per QALY respectively. The threshold analysis showed that as the annual cost of DAART increases, the cost of this treatment becomes less and less cost-effective (Fig 3). For DAART to be cost-effective its annual cost needs to be below US$500.

The output for the probabilistic sensitivity analysis for 10,000 simulations is shown in Fig 4. All of the model outputs were in the northeast quadrant of the cost-effectiveness plane, suggesting that DAART is more costly and more effective than self-administered ART, it is never cost saving, and never has a negative impact on patient outcomes. At a threshold of US$9,000, DAART was found to be 50% likely to be cost-effective, and if the willingness to pay for a QALY was US$18,000 then DAART is likely to be at least 95% cost-effective. The probability that DAART was cost-effective at the WHO-CHOICE threshold of US$5,086 was just 1% (Fig 5).

The population EVPI is illustrated in Fig 6. At cost-effectiveness threshold of $5,086, the population EVPI becomes substantial (≈US$492 million), and is likely to exceed the cost of additional investigation. This suggests that further research will be potentially cost-effective. The EVPI for each group of model parameters is illustrated in Fig 7 for a threshold for cost-effectiveness of US$5,086. The value of information associated with research on the cost of DAART and its effectiveness is US$1,706,707 and US$853,353 respectively. The other groups of model inputs (see Table 1), such as transitional probabilities, QALYs gained and cost of treatment, have no value of information associated with them. This suggests that if further research is commissioned, it should focus on the relative effectiveness of DAART compared to self-administered ART and its cost.

The cost-effectiveness of DAART versus self-administered ART was analysed using a model based cost-utility analysis. The ICER of $8,700 per QALY gained indicates that when considering the WHO-CHOICE threshold of US$5,086, from the perspective of health care payer in sub-Saharan Africa, directly administered ART cannot be regarded as cost-effective. The threshold analysis conducted for a range of costs for DAART, showed that for it to be cost-effective, the annual cost of DAART needs to be priced below $500 per patient. The results from the one-way sensitivity analysis indicate that the ICER is only sensitive to the estimates of the effectiveness and annual cost of DAART. The value of information analysis showed that further primary research will be worthwhile and potentially cost-effective in resolving the uncertainty on whether or not to adopt DAART among HIV patients at high risk of non-adherence. This should focus on establishing the relative effectiveness and cost of DAART.

To the authors’ knowledge there are no comparative studies that have examined the cost-effectiveness of DAART in a sub-Saharan Africa setting. This reflects a wider paucity of economic evidence around HIV HIV treatment in low- and middle-income countries [35]. In a closely related study that had a different patient population, McCabe and colleagues constructed a mathematical model to examine the cost-effectiveness of DAART versus self-administered ART among pregnant HIV-infected women in their third trimester in the US? [36]. Their study found that the DAART was cost-effective from the perspective of health care payer in the United States (ICER $14,233 per QALY) [36].

We used a model based approach [37] which had several advantages [37, 38]. It allowed the extrapolation of the outcomes reported in the literature to reflect long term consequences and thus capture important economic outcomes. The effectiveness data were extracted from a meta-analysis of multiple clinical trials instead of using data from just one clinical trial [5]. Another important strength is that the cost data were taken from recent studies. There are some limitations of this study that must be acknowledged. The main limitation pertains to the inherent uncertainty in model input parameters and the observed sensitivity of the ICER to certain key parameters. For example, some caution is required around the estimates of the cost of DAART and self-administered ART as this may vary depending on the particular health care systems in different countries. However, several sensitivity analyses were conducted to assess both the model and structural uncertainties, including one-way sensitivity analysis, threshold analysis and probabilistic sensitivity analysis. The ICER results are robust over a range of model parameters and varying most of the parameters from worst-case to best-case scenarios did not significantly impact on the results. Another important limitation is that the treatment effect estimate used for modelling DAART effectiveness as based on a meta-analysis that included trials from heterogeneous contexts (prison, methadone clinics), and only a minority of the included studies were conducted in sub-Saharan Africa. Additionally, we used evidence from the Hart meta-analysis[5] that included only randomised controlled trials which found no evidence of DAART effectiveness over the Ford meta-analysis[15] that included both the randomised controlled trials and non-randomised controlled trials.

In addition, the model adopted a simplistic representation, we did not consider the impact of DAART on transmission of HIV in the target population. The main limiting assumption of such approach includes the use of non-dynamic process at the level of the risk of each sexual encounter, which could lead to a large underestimate of benefit. Similarly, the model could underestimate cost effectiveness of DAART for prevention of HIV, because we assume no direct benefit to the female partners. We focused only on DALYs averted.