We analysed the direct health care costs of treating over 10,000 HIV-infected adults enrolled in a Southern African managed care ART programme with almost 600,000 patient months of follow-up, spanning 3 y before ART to 5 y on ART. We found a peak in costs in the period around the time of ART initiation, thereafter total mean costs dropped off to a plateau that persisted for 5 y. An important and novel feature of our study was the presentation of time-dependent associations between total mean costs and relevant variables. We identified lower baseline CD4+ cell count, higher baseline viral load, and shorter duration of CD4+ cell count monitoring before starting ART (as a proxy for HIV care) as being independently associated with higher costs in the early time periods. Lower ART adherence, being on second line ART, and starting ART at an younger age were most strongly associated with lower mean costs in later time periods, and the association with ART adherence became more marked over time.
The peak in costs in the peri-ART period we observed was largely driven by the high proportion of patients requiring hospitalisation. High rates of early morbidity, often resulting in hospitalisation or death, are characteristic of antiretroviral programmes in resource-limited settings. Patients on ART in low-income countries have higher early mortality compared with high-income countries, even after correcting for baseline differences in CD4+ cell counts 
. A strength of our study is the analysis of cost data before starting ART. Few ART cost analyses include the period before starting ART. Higher costs in the first year on ART compared with later years with high rates of hospitalisation was reported in another South African study of a public sector ART programme, but they only assessed costs for 1 mo before starting ART and did not attempt to more accurately define the period of high cost 
. Given our finding of high costs in the 4-mo period before starting ART, which was equivalent to 1.5 y of cost in patients on ART after the first year, other studies might have significantly underestimated the costs of providing HIV care just prior to starting ART.
We found that higher ART adherence was associated with lower costs particularly after removing antiretroviral drug costs. The magnitude of this association becomes greater as duration on ART increases. However, the continuous model showed that while highly adherent patients (>92%) were associated with the lowest total mean costs in later time intervals, they were associated with higher costs in the early time intervals. A similar association was found with high baseline CD4+ cell counts (>300 cells/µl) being associated with higher costs initially. These findings could be attributed to increased health-seeking behaviour leading to increased costs initially, but reduced costs over time. Very low ART adherence was associated with low total mean costs in all time intervals as these patients are presumably accessing minimal services. Our group has previously reported that ART adherence assessed by pharmacy refills in this cohort predicted both virological suppression 
and survival 
. Poor adherence limits the effectiveness of ART, drives resistance to first line regimens, and thus leads to earlier switching to costly second line ART. Despite the important role of ART adherence, existing economic models fail to include it.
Our analysis of the time-dependent associations with increased costs has several important public health implications. The high early costs of ART programmes could be reduced by starting ART at a CD4+ cell count of <350 cells/µl rather than <200 cells/µl (for patients without major symptomatic HIV disease). Our cohort does not allow for an evaluation of starting ART in patients with baseline CD4+ cell counts ≥350 cells/µl because AfA guidelines only allow these patients to start ART following an AIDS-defining illness or with other serious co-morbidity: costs were actually higher in this group compared with those starting ART with baseline CD4+ cell counts 200–349 cells/µl, presumably reflecting the costs of treating the morbidity that was the criterion for starting ART. The second intervention that could reduce early costs would be the earlier identification of HIV infection, illustrated by our finding that being in HIV care for more than 6 mo prior to starting ART reduced costs in the peri-ART period. The key driver of later costs with public health implications is ART adherence. Higher adherence prolongs time on the cheaper first line regimen, but also reduces non-ART direct costs in our study. The third intervention that might reduce costs would be to encourage ART programmes to invest in systems to monitor ART adherence and implement effective interventions if adherence is suboptimal. ART adherence could be monitored over short time periods of 3 to 4 mo, which identifies patients incurring higher costs and those at risk of virological failure 
We estimate that annual total direct health care costs are approximately US$2,400 (after the peak in costs in the peri-ART period) for patients accessing ART in the private sector. Lower costs were reported in two other South African studies. Harling reported costs of $2,502 in year one and $1,372 in year two of a donor-funded public sector program 
. Rosen estimated the ART component of care to be US$757–US$1,126 in the first year of several different models of ART delivery to public sector patients, but non-ART–related clinic visits and hospitalisations were not included 
. The incidence rate of hospitalisation we found in the first 6 mo on ART was similar to that reported in a South African public sector ART programme in the first 48 wk on ART, but our incidence was higher in later periods, which would increase costs 
. Higher rates of hospitalisation in the private sector compared with the public sector after the initial period of ART probably reflect greater access in the private sector. Other factors driving higher costs in the private sector compared with public sector are higher costs for hospitalisation (US$340 versus US$202 per day, respectively) and viral load tests (US$62 versus US$42, respectively) 
Some findings of other ART programme cost studies differed from our analysis. We found that the ART component of costs was relatively small compared with other studies in resource-limited settings 
, which could be related to higher hospitalisation and other costs in our private sector setting. Younger age has been found to be associated with increased costs in some 
, but not all studies 
. We found a significant age effect with younger age (<25 y) associated with lower early but higher later costs and older age (≥50 y) associated with higher early and especially later costs. Finally, unlike the finding of another South African study 
, sex was not independently associated with costs, even after controlling for pregnancy-related costs and the higher proportion of men being on efavirenz. It is possible that our inclusion of ART adherence in our multiple regression model adjusted for sex differences, as we have previously shown that men have lower ART adherence than women 
There are a number of limitations to this analysis. First, our cohort consisted of private sector patients when the majority of patients in resource-limited settings are treated in the public sector. However, the baseline characteristics of our cohort (CD4+ cell count, proportion of females, and age) are comparable with cohorts from low-income countries 
. The BMI was <18.5 kg/m2
in 13% of our cohort compared with 19% in a South African public sector cohort 
, but their patients had more advanced disease as evidenced by their lower baseline CD4+ cell counts. These baseline nutritional differences would likely impact outcomes. We restricted our analysis to patients receiving NNRTI-based first line ART regimens, in keeping with WHO recommendations for resource-limited settings 
. While we would not claim that our actual cost findings are generalisable to public sector settings or to other countries, we would argue that the variables that drive early and late costs are likely to be relevant even if the magnitude of the effect could differ.
Second, the impact of specific AIDS-defining illnesses on outcomes and costs was not included in this analysis because these data were not available. Third, as a provider's perspective was chosen for this analysis, the cost to society is not fully represented because we did not have data on direct non-health care costs and indirect costs. However, a provider's perspective is more appropriate for the aim of this study, which was to unpack the key drivers of health care costs in order to inform appropriate budgeting and planning. Fourth, the characteristics of the patients who left the scheme were different from those who remained, which may have affected our findings. However, there was no significant difference in the key baseline characteristic of CD4+ cell count and many of the other differences (e.g., age, difference of 0.1 log10
viral load) were small and are of questionable importance. Fifth, we chose to use the tariff amount as opposed to the amount claimed or reimbursed so that similar services would take the same monetary value and have further assumed that these tariffs are a suitable proxy for opportunity costs. While this could be a shortcoming, it is common to assume that market prices are a proxy for opportunity costs in economic evaluation given the difficulties in evaluating the latter 
. Finally, cost minimisation should not be the only goal of health care providers, and other important aspects of care such as quality and outcomes are not addressed by our analysis.
In conclusion, we have described the temporal trends of costs of a large private sector HIV disease management programme in Southern Africa and shown that associations with costs change over time. Interventions that should reduce early costs include starting ART at higher CD4 counts and being in HIV care for longer periods before starting ART. Our results also indicate that systems to detect suboptimal ART adherence and interventions that improve adherence would reduce later costs. The increasing impact of ART adherence on costs over time suggests that this variable should be incorporated in economic models of ART.