https://orcid.org/0000-0003-2179-0251
Figures
Abstract
Background
Replacing conventional, facility-based HIV treatment with less intensive differentiated service delivery (DSD) models could benefit DSD clients and the health system, but its value depends on maintaining or improving clinical outcomes. We compared retention and viral suppression between antiretroviral therapy (ART) clients enrolled in DSD models to those eligible for but not enrolled in DSD models in South Africa.
Methods and findings
We applied a target trial emulation (TTE) methodology to data from South Africa’s electronic medical record system (TIER.Net) for 24 public-sector health facilities across three provinces and estimated retention in care (attended facility visit within 12 months) and viral suppression (<400 copies/ml3) at 12, 24, and 36 months after follow-up start date, defined as DSD enrollment date for the intervention arm and the first trial enrollment period facility visit for the comparison arm. Clients were eligible for DSD models if they were ≥18 years old, on ART ≥12 months, and had two suppressed viral load (VL) measurements, per prevailing national guidelines. For the TTE, we designated eight 6-month target trial enrollment periods between 1 July 2017 and 1 July 2021. For each period, we estimated the risk differences for retention in care and viral suppression by comparing those enrolled in DSD models to those not enrolled, using a Poisson distribution with an identity link function. We report adjusted and unadjusted risk differences for clients enrolled in DSD models and for DSD-eligible clients not enrolled in a DSD model. Estimates were adjusted for age, sex, urban/rural facility setting, province, WHO stage at ART initiation, and years on ART at trial enrollment.
49,595 unique individuals were eligible for DSD enrollment over eight target trials, contributing to a total of 148,943 trial-clients, of whom 17% (25,775) were enrolled in DSD models. The pooled adjusted risk difference for retention in care between clients enrolled in DSD and those not enrolled in DSD was 3.2% (95% confidence interval (CI) [1.6%,4.7%]) at 12 months, 4.2% (95% CI [2.4%,6.0%]) at 24 months, and 4.4% (95% CI [2.0%,6.8%]) at 36 months. For viral suppression, the adjusted risk difference comparing DSD to non-DSD was estimated to be 1.4% (95% CI [−0.5%,3.2%]) at 12 months, 1.7% (95% CI [−0.5%,4.0%]) at 24 months, and 1.4% (95% CI [−0.6%,4.4%]) at 36 months. Results remained consistent across target trials. Clients who were younger, received care from a facility in an urban settings, or had less ART experience at trial enrollment had lower retention. Study limitations include reliance on routinely collected medical records and the likely presence of residual confounding.
Conclusions
Clients enrolled in DSD models in South Africa had slightly better retention in care and similar viral suppression to those who were eligible for but not enrolled in DSD. With better or equivalent outcomes, DSD models can be assessed on the basis of non-clinic costs and benefits, such as changes in quality of care and resource utilization.
Author summary
Why was this study done?
- Differentiated service delivery (DSD) models are a client-centered approach to tailor the timing, location, provider, and type of HIV service delivery to address the needs of people living with HIV while reducing pressure on the healthcare system.
- In South Africa, where this work was done, the main models of care are facility and external pick-up points, and to a lesser extent, adherence clubs. These models dispense 2–3 months of medication a time, and clients see a nurse every 6 months for a clinical review.
- One expectation of DSD models is that both clients and the health system will benefit or at least have similar outcomes. Clients eligible for these models are established on care and virally suppressed at enrollment, so the aim is to maintain viral suppression.
What did the researchers do and find?
- We used electronic medical record data (TIER.Net) to measure outcomes of clients who were enrolled in a DSD model and then compared them to outcomes for clients who were not enrolled in a DSD model.
- We utilized the target trial emulation framework to assist in estimating causal effects from observational data. When using this method, researchers first specify the protocol of the hypothetical randomized trial, then use observational data to emulate its components.
- We then estimated the risk differences for retention in care and viral suppression comparing those in DSD to those not enrolled in DSD models (but eligible), finding that clients enrolled in DSD models in South Africa had slightly better retention in care and similar viral suppression to those who were eligible for but not enrolled in DSD.
What do these findings mean?
- Overall, we observed that these established clients were doing well. While there might be a slight benefit to being in DSD, we found no evidence to suggest any evidence of harm.
- The target trial emulation methodology cannot correct all issues that affect observational studies, and residual confounding can still be present.
- Bias is very likely with regard to who was offered DSD enrollment, particularly during the initial phase of the DSD roll out due to the expectations of both providers and clients about the future.
- Further work to explore the systems, patient, and provider-level impacts of DSD, particularly the resource utilization and costing analysis need to be conducted to understand the full impact of these interventions.
Citation: Huber AN, Jamieson L, Fox MP, Manganye M, Malala L, Chidarikire T, et al. (2025) Evaluating the impact of differentiated service delivery (DSD) on retention in care and HIV viral suppression in South Africa: A target trial emulation using routine healthcare data. PLoS Med 22(8): e1004489. https://doi.org/10.1371/journal.pmed.1004489
Academic Editor: Elvin Hsing Geng, Washington University in St Louis School of Medicine, UNITED STATES OF AMERICA
Received: October 2, 2024; Accepted: July 1, 2025; Published: August 26, 2025
Copyright: © 2025 Huber et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data used in this study are owned by the South African National Department of Health and cannot be shared by the authors. South African The National Health Research Database website (https://nhrd.health.gov.za/) provides information on the application process for research and data requests that fall under the National Health Research Committee and the 9 Provincial Health Research Committees in South Africa.
Funding: Funding for the study was provided by the Bill & Melinda Gates Foundation (https://www.gatesfoundation.org/) through OPP1192640 to Boston University (SR as principal investigator) and INV-037138 to the Wits Health Consortium (SP as principal investigator). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Competing interests: I have read the journal's policy and the authors of this manuscript have the following competing interests: LM, MM and NK are employed by the government agency that has supervisory authority over the healthcare facilities from which the data for this study were drawn. SR is a member of PLOS Medicine's editorial board.
Abbreviations: ART, antiretroviral therapy; CCMDD, Central Chronic Medicine Dispensing and Distribution; CDU, Central Dispensing Unit; CI, confidence interval; DSD, differentiated service delivery; IQR, interquartile range; SSA, sub-Saharan Africa; TTE, target trial emulation; VL, viral load; WHO, World Health Organization
Introduction
In South Africa, as in other countries in sub-Saharan Africa (SSA), the use of differentiated service delivery (DSD) models for HIV treatment is expanding rapidly [1,2]. For eligible antiretroviral therapy (ART) clients, DSD models adapt the characteristics of service delivery, such as location and frequency of healthcare system interactions, to meet the needs of different client groups. This approach is expected to sustain clients’ positive clinical outcomes, increase satisfaction with their care, and increase the efficiency of healthcare systems by decongesting clinics, reducing pressure on healthcare providers, and simplifying procedures [3,4]. Most DSD models are limited to “established” or stable ART clients who have been on treatment for at least 6 months, have documented viral load (VL) suppression, and do not have uncontrolled co-morbid conditions or HIV-related symptoms [5]. The proportion of established clients enrolled in any DSD model, rather than remaining in conventional (non-differentiated) care, varies widely by country and facility, based on guidelines, resources, and provider and client preferences, among other factors [6].
South Africa, home to the world’s largest HIV treatment program [7], offers three main less intensive models for ART clients who are established on treatment and have documented viral suppression: facility-based medication pickup points, external (out-of-facility) medication pickup points, and adherence clubs. (In South Africa, these models are referred to as Differentiated Models of Care, or DMOC and fall within the category of Repeat Prescription Collection strategies or RPCs.) The three models for established ART clients are supported by the Central Chronic Medicine Dispensing and Distribution (CCMDD) programme, the Central Dispensing Unit (CDU), or facility-based pharmacies. They provide centralized medication packaging and delivery to clients, with options for pickup at either facility pick-up points within healthcare facilities, or at external pick-up points adjacent to facilities or within communities [7–10]. Other DSD models, such as medication lockers and home delivery, have also been implemented in South Africa, but on a much smaller scale.
Several randomized trials and observational studies have reported the outcomes of clients enrolled in South Africa’s less intensive models [11–13] and in others outside South Africa [14] and found high rates of retention in care and viral suppression after one or more years of model exposure, an unsurprising result given that retention and suppression are also both eligibility criteria for model enrollment. Few studies [15–17], however, have compared the outcomes of clients actively enrolled in DSD models to those who are eligible for DSD models but remain in conventional care. It is these eligible-but-not-enrolled clients who comprise the relevant comparison group.
The absence of evidence about how clients’ outcomes in DSD models compare to those of similar clients remaining in conventional care hampers decisions about whether to expand DSD programs and how to update eligibility criteria. To help fill that gap, we assessed the effectiveness of DSD by comparing clients on DSD to those eligible for but not on DSD on outcomes of retention in care and HIV viral suppression at 12, 24, and 36 months using routinely collected electronic medical record data.
Methods
Using a target trial emulation (TTE) approach, we analyzed routinely collected medical records from a sample of public-sector health facilities. We compared the 12-, 24-, and 36-month retention in care and viral suppression outcomes of DSD clients to those eligible but not enrolled in DSD in South Africa at public-sector health facilities between 1 July 2017 and 1 July 2021.
Models of care
As noted above, DSD-eligible ART clients were either enrolled in one of three less intensive differentiated models of care in use during our analysis—primarily facility-based medication pickup and external medication pickup, along with adherence clubs for a smaller number of clients—or in conventional (non-differentiated) care. Enrollment decisions were presumably based on client eligibility, client choice, provider preference, the availability of each model at any specific healthcare facility, and the facility’s access to enrollment procedures (e.g., the CCMDD database). Features of the four models represented in our analysis are described in Table 1. Though TIER.Net contained a field for capturing DSD enrollment during the study period, it was not filled out for most clients. We were thus not able to reliably distinguish among the three DSD models in the TIER.Net data and could only evaluate the impact of all DSD models compared to conventional care, rather than comparing DSD models with one another. The 2016 South African Adherence Guidelines were used to define DSD eligibility and viral suppression in this analysis, as the majority of the enrollment period fell within their period of coverage [18]. During the study period, eligibility for DSD enrollment required clients to be non-pregnant adults with a minimum of 12 months of experience on ART and at least two documented suppressed VLs (<400 copies/ml). We note that under current guidelines, the VL threshold for suppression is 50 c/mL and 3-month dispensing is recommended.
Study sites and population
Data were drawn from 24 public-sector health facilities in four districts in three provinces of South Africa (6 facilities in each district—West Rand and Ekurhuleni in Gauteng Province, King Cetshwayo in KwaZulu-Natal Province, and Ehlanzeni in Mpumalanga Province). We purposively chose a mix of high- and medium-volume facilities in urban and rural settings. All sites included offered at least one DSD model at the time of data collection had at least 1,000 clients on ART and utilized South Africa’s electronic medical record system, TIER.Net [19]. The data extracted from the study sites spanned the period from July 2017 to June 2023, depending on the district, and in 2021 the approximate proportion of ART clients enrolled in DSD models ranged between 24% and 81%, with KwaZulu-Natal reporting the highest proportion of clients enrolled. (S1 Table).
The TIER.Net system [20] is a facility-based electronic register. Each client’s health information is recorded in a paper file at the point of care and then entered into TIER.Net by an on-site data capturer. Data are collected prospectively, and inactive records are not deleted. The system allows for capture of age, sex, HIV test date, ART start date, visit date, ART medication prescribed, and months of ART dispensed for all facility interactions, date and result of HIV-related laboratory tests, and the outcome date and status for clients who have died, transferred, or become disengaged from care. The sex variable was captured at ART initiation by the healthcare worker and was likely based upon either self-report or a clinical examination, if conducted. The outcome variable may include information from telephone or home visit tracing if the facility provides these services. A DSD model enrollment variable was introduced in 2016, but with three limitations for DSD program monitoring. First, out-of-facility medication collection from external pickup points and community-based adherence clubs is not captured in TIER.Net—only actual facility visits are recorded. Second, before June 2020, the DSD enrollment variable did not distinguish among separate DSD models, but only the client’s enrollment in any non-conventional model. And third, unofficial (silent) transfers of clients between facilities are not captured in TIER.Net, creating a potentially large discrepancy between facility-level and overall ART program loss to follow-up [21]. To address these limitations, we grouped all DSD models together included a 12-month window for follow-up visit measurement, and reported facility-level retention.
Target trial emulation
TTE is a methodology that offers researchers a framework for analyzing observational data in a way that simulates a hypothetical randomized clinical trial. Its goal is to estimate causal effects while reducing biases commonly found in observational studies [22]. It also helps to prevent immortal time bias, whereby a person can accrue person-time in the study when they cannot logically experience the outcome. Immortal time bias often occurs when the beginning of follow-up time and intervention exposure are not aligned [23]. Using this methodology, we clearly described the hypothetical trial we would like to have conducted and then attempted to mimic that trial as closely as possible using observational data, potentially reducing bias in our assessment of the effect of DSD on treatment outcomes.
Table 2 outlines our hypothetical target trial and the choices we made in our observational study to best emulate that trial. Both the target trial and the emulation sought to estimate the impact of enrollment in DSD models on retention in care and VL suppression at 12, 24, and 36 months after DSD eligibility. To do this, we used in both the target trial and the emulation the DSD eligibility criteria in the South African guidelines [21] in effect at the time of the study [18]. In addition, we included in the emulation additional trial eligibility criteria in line with the conduct of our hypothetical trial: (1) documentation of a recent facility visit, within 6 months before the trial, and (2) no prior DSD enrollment recorded (DSD naïve at study enrollment). For this analysis, there was no prospective study protocol or analysis plan detailing the analyses presented.
Our target trial intervention was enrollment in a DSD model. After target trial randomization, participants could return to standard care at the discretion of the client and the provider. We mimicked this in our emulation by defining the intervention to be enrollment in DSD at baseline. After DSD enrollment, participants could return to regular standard care, but we would still analyze them as if they were in the DSD arm as per intention-to-treat, as would have been done in the target trial. The comparison arm for our target trial was the participants randomized to not immediately enroll in a DSD model. As would happen in routine care, however, comparison arm participants would be allowed to enroll in a DSD model at a later time point after the enrollment period, should they be offered the opportunity and accept it. This effectively alters our study question to comparing the effect of enrolling in DSD immediately to the effect of a delayed opportunity for DSD enrollment, with some comparison arm participants later enrolling in DSD and some never doing so. To mimic this in our emulated trial, we assigned anyone within the baseline enrollment period who did not enroll in DSD to the comparison arm, noting that after baseline, they had regular standard care, which in some cases could involve later enrolling in DSD.
Development of multiple trial emulations
For this analysis, we chose to emulate a series of eight target trials, each encompassing a 6-calendar-month DSD enrollment period in our data set as our baseline, between 1 July 2017 and 1 Jul 2021 (e.g., 1 July 2017 to 31 December 2017, 1 January 2018 to 30 June 2018, etc.). This approach maximized the value of our multi-year data set within the constraints of the target trial methodology. Had we conducted only one emulated trial limited to the first 6-month calendar period for which we had data (July to December 2017), all participants could readily have been assigned to either intervention or comparison arm status, but this approach would have excluded all the individuals in the data set who enrolled in or became eligible for DSD enrollment after December 2017. Alternatively, we could have conducted only one emulated trial incorporating all the data in our data set (July 2017 to June 2021), in which all clients who enrolled in DSD later in the study period would have been assigned to the DSD arm. As those clients would, by definition, have survived longer on ART before enrolling in DSD, however, and longer duration on ART is associated with better retention and suppression outcomes, this could create bias.
Emulating multiple, time-limited trials offers one solution to the problem described above. Within each of the eight target trial periods, we assessed eligibility for participation in the trial and DSD models for individuals in the database at the start of each 6-month trial (Fig 1). All those who met the criteria were included in the new trial, and individual clients could thus be included in the comparison arm of multiple trials. Participants eligible for each 6-month calendar period were then assigned to the intervention or comparison arm. Because being DSD-naïve is a trial eligibility criterion, anyone who had enrolled in DSD during a prior 6-month period and been analyzed in a previous emulated trial would no longer be eligible for subsequent trials. A participant who had not yet enrolled in DSD in a prior period could meet the study eligibility criteria for subsequent trials, however, along with anyone who became newly eligible for DSD based on national guidelines. As a result, any individual could be enrolled in one or more trials as a comparison arm participant only; in one or more trials as a comparison arm participant and one as an intervention participant; or in just one trial as an intervention participant.
DSD, differentiated service delivery; FU, follow up.
Outcomes
We defined two outcomes for this analysis, retention in care and viral suppression, each assessed at 12, 24, and 36 months after the start date of follow-up as specified in Table 2. Retention in care was defined as having had a recorded facility visit at the same public-sector health facility at 12, 24, or 36 months after the start date, with the start date defined as the DSD enrollment date for the intervention arm and the first facility visit attended during the trial enrollment period for the comparison arm. Viral suppression (<400 copies/mL) was also assessed at these time points but could only be reported for those with VLs measured (i.e., our denominator only includes those who had documented VL results in their medical records); participants missing VL measures were excluded from the analysis of this second outcome. It is important to note that because our data source, TIER.Net, does not capture unreported (“silent”) transfers between facilities, those who informally transfer appear to have disengaged from care. Retention in this study thus refers to retention at the same facility, not retention across the HIV treatment program.
For both retention in care and viral suppression outcomes, we allowed for a wide 12-month window after the endpoint to recognize outcomes. Clients were considered retained in care for 12 months, for example, if they had a facility visit any time between 12 and 24 months after DSD enrollment, and their VL outcome for the 12-month endpoint reflected laboratory tests conducted between 12 and 24 months after DSD enrollment. These relatively long (12-month) windows were chosen during the analysis phase to accommodate local practice during the study period, which included the COVID-19 pandemic when South Africa temporarily allowed 12-month prescriptions of ART medications. A 12-month window for recording viral suppression was necessary because VL tests are recommended and typically conducted only every 12 months after the first year on ART. We used the earliest VL result within the 12-month window after the specified endpoint as the outcome measure. We also extracted from TIER.Net individual characteristics such as age, sex, time on ART, World Health Organization (WHO) stage, and CD4 count at ART initiation and utilized the TIER.Net outcome variable to classify those who were not retained as dead, transferred, or lost to follow-up.
Follow-up for all clients continued for up to 48 months from the start date through their electronic medical records in the TIER.Net system. We applied administrative censoring for each 12-, 24-, and 36-month outcome to ensure clients had sufficient potential follow-up data available to assess outcomes. Clients were censored if the outer window for measuring each outcome went beyond the data extraction date. For assessing the 12-month outcome, for example, clients included in the analysis had to have at least 24 months of potential follow-up data available (= 12-month follow-up period + 12-month window for outcomes). Clients starting in 2021 or later had not accrued enough follow-up time for the 24- and 36-month outcomes. These clients were excluded from having their outcomes assessed at 24 and/or 36 months and are denoted as having “insufficient follow-up” for those time points.
Statistical analysis
We first describe the characteristics of the unique clients in the analytic cohort in the study. Since individual participants could be included in multiple trials, however, our unit of analysis was a trial-client, rather than a unique client. Trial clients are also described (S1 Fig). Once all eight target trials were defined and the cohorts were created, we estimated the risk differences for retention in care and viral suppression comparing those in DSD to those not enrolled in DSD models using a Poisson distribution with an identity link function. We adjusted for age, sex, urban/rural health facility setting, province, WHO stage at ART initiation, and years on ART at trial enrollment. We did not adjust for CD4 count at ART initiation for two reasons: (1) potential collinearity with WHO stage and (2) CD4 count had more missing data compared to WHO stage. We did conduct a sensitivity analysis in which we adjusted for CD4 count at ART initiation in place of WHO stage, and these results are reported in S4 Table. We also conducted an age-stratified and sex-stratified analysis to further evaluate the age- and sex-specific impact of DSD on health outcomes and conducted smaller sub-analyses to describe the TIER.Net reported outcomes of those not retained within the given outcome window period. We present risk differences and 95% confidence intervals (CIs) for each of the eight trials and a pooled estimate that further adjusts for the individual target trial. To account for clients appearing in multiple target trials (resulting in multiple trials for some that would artificially inflate the sample size and bias our standard errors), we also adjusted for within-participant variance using robust variance estimates.
Results
Analytic cohort
A total of 49,595 unique individuals were eligible for DSD model enrollment in our data set. Of these, 52% (25,775) were enrolled in DSD models during the observation period, while 23,820 were eligible but never enrolled. The distribution of sex and WHO stage at ART initiation was similar between those ever enrolled in DSD and those who remained in standard care, with two-thirds (67%) of the clients in WHO stage 1 at baseline and most (69%) female (Table 3). The age and median time on ART at the start of the first trial were higher for those enrolled in DSD (51% aged 35–49 years; median time on ART 4.0 years [interquartile range (IQR) 2.3 to 6.5 years]) compared to those not enrolled in DSD (42% aged 35–49 years; median time on ART 2.5 years [1.5 to 5.3 years]). Clients in the youngest age group (18–24) were less likely to be enrolled in DSD than were older clients. Clients ever enrolled in DSD were slightly more likely to receive care in urban settings than were those never enrolled in DSD (75% versus 69%, respectively) (Table 3). Cohort characteristics based on client trials, rather than unique individuals (available in S2 Table), are more balanced between DSD and our comparator arms.
As explained above, because individual participants were included in multiple trials, our unit of analysis was a trial-client, rather than a unique client. 580,276 trial clients were assessed over the eight target trial enrollment periods between July 2017 and July 2021 (Fig 2). Of all trial clients, 431,333 were excluded for different reasons: a third (34%, n = 200,125) were previously exposed to DSD models, 5% (n = 27,791) had not had a recent facility visit before trial enrollment, 3% (n = 18,695) were aged <18 years, 2% (n = 9,623) were pregnant during the trial enrollment period, 16% (n = 69,136) had <12 months of ART experience, and 25% (n = 105,963) were not virally suppressed(Fig 2). A detailed disaggregated flow diagram depicting each of the eight target trials is available in S1 Fig.
DSD, differentiated service delivery.
Outcomes
The adjusted mean proportions of participants retained at 12-, 24-, and 36-month post-enrollment period were 92%, 87%, and 82% in the DSD arm, respectively, while proportions retained in the non-DSD comparison arm were 89%, 83%, and 78%, respectively (Fig 3A). The estimated adjusted mean proportions virally suppressed, among those who had VL results in their medical records, were 95% at all time periods for those enrolled in DSD models and 93%−94% at all time periods for those not enrolled in DSD models (Fig 3B).
Estimates adjusted for age, sex, urban/rural facility setting, province, WHO stage at ART initiation, years on ART at trial enrollment.
The pooled adjusted risk difference for retention in care between clients enrolled in DSD and those eligible but not enrolled was estimated to be 3.2% (95% CI [1.6%,4.7%]) at 12 months, 4.2% (95% CI [2.4%,6.0%]) at 24 months, and 4.4% (95% CI [2.0%,6.8%]) at 36 months (Fig 4). Estimates stratified by each of the target trials were similar, except for the 12-month outcome in the 7th trial which could be explained by timing of the trial enrollment period, which occurred during the first two waves of the COVID-19 pandemic.
Clear circles, unadjusted estimates; black circles, adjusted estimates; error bars indicate 95% confidence intervals; estimates adjusted for age, sex, urban/rural facility setting, province, WHO stage at ART initiation, years on ART at trial enrollment.
For viral suppression outcomes, the adjusted risk difference comparing DSD to non-DSD was estimated to be 1.4% (95% CI [−0.5%,3.2%]) at 12 months, 1.7% (95% CI [−0.5%,4.0%]) at 24 months, and 1.4% (95% CI [−0.6%,4.4%]) at 36 months (Fig 5). Between 82%–85% of people retained at each time point had a VL measurement; there was no major difference between the two groups in VL data availability. The pooled adjusted risk differences, including the covariate estimates are available in S3 Table. These show that clients who were younger (aged 18–34 years), received care from an urban facility, or had fewer years on ART (<2 years) at study enrollment were significantly less likely to be retained in care, while those aged 18–24 years had lower rates of viral suppression at 12 months (S3 Table). Adjusting for CD4 count at ART initiation, instead of WHO stage, did not fundamentally change the effect estimates for those on DSD models (S4 Table).
Clear circles, unadjusted estimates; black circles, adjusted estimates; error bars indicate 95% confidence intervals; estimates adjusted for age, sex, urban/rural facility setting, province, WHO stage at ART initiation, years on ART at trial enrollment.
Age-stratified analyses showed some variation in risk difference estimates between age groups when comparing DSD versus non-DSD rates of retention and viral suppression. In particular, younger participants (aged 18−24 years) had a higher risk difference for both retention and viral suppression compared to those older (aged 25+ years) (S5 and S6 Tables). Adjusted risk differences for retention ranged from 4.4% to 7.8% for 18–24-year-old clients and from 2.5% to 5.4% for clients aged 25+ years, while for viral suppression they ranged from 3.3% to 6.1% (aged 18−24 years) and from 0.9% to 1.7% (aged 25+ years). CIs were wider for 18–24-year-old clients due to the smaller sample sizes involved, however. Sex-stratified analyses produced similar risk difference estimates between males and females for retention. There were some differences in viral suppression by sex, however, with male clients having a slightly higher adjusted risk difference estimate comparing DSD versus non-DSD (ranging from 1.8% to 2.2%) than did female clients (range 1.1%−1.5%) (S7 and S8 Tables).
In additional sub-analyses, we found that a small portion of all target trial patients had either died during the outcome window period (0.2%–0.4% at 12, 24, and 36 months) or had transferred to another facility (1%–2%). These findings were similar between the DSD and non-DSD arms (S9 Table).
Discussion
In this study, we conducted an analysis of a large cohort of clients comparing viral suppression and retention outcomes between those enrolled in DSD models and those eligible for DSD enrollment but remaining in conventional care in a routine healthcare setting. We found that clients enrolled in DSD models had similar or slightly better rates of retention in care and VL suppression (82%−92% and 95%, respectively) as did those who were eligible for DSD but remained in conventional care (78%−89% and 93%–94%, respectively).
For clients eligible for DSD—i.e., those considered “established” on ART—it is reassuring that retention and suppression outcomes were comparable between conventional care and DSD models. There was little room for these clients’ rate of viral suppression to improve over the follow-up period, as all had documented viral suppression at enrollment. The concern that less intensive models of service delivery, with less direct interaction with clients than under conventional care, might lead to worse retention in care, however, was not borne out in our analysis. The potential benefits of DSD, in this setting, can thus be assessed based on non-clinical factors, such as costs to clients, health system resource utilization and efficiency, and client and provider satisfaction.
While countries across SSA have implemented differentiated models of HIV treatment, few of these programs have undergone rigorous evaluation at scale. Many studies have focused on the impact of individual interventions, often in controlled settings, while less attention has been given to reporting outcomes of these DSD models in routine care settings after national implementation. Despite differences in study designs, our results largely align with the majority of studies on individual models, indicating that less intensive DSD models enrolling clients already established on ART provide equivalent or slightly better retention and viral suppression outcomes [11–17]. Our findings are also consistent with a recent large-scale study assessing the impact of DSD models on 12-month retention in Mozambique [17]. In that study, the research team conducted an interrupted time series analysis comparing facility-wide retention before and 12 months after the implementation of a package comprising eight models of care. We note that while the question of facility-wide retention addressed in that analysis differs from the one addressed in our paper, which specifically focuses on the eligible population, both perspectives merit consideration.
Our study had several limitations, most resulting from the observational design and reliance on routinely collected medical records. Although the use of TIER.Net data provided a large sample of ART clients attending multiple healthcare facilities, until June 2020 it did not record the specific DSD model in which a client was enrolled and instead only captured whether the client was in a DSD model or not. All of the models of care in which participants were enrolled had the same schedule of clinical consultations and medication dispensing intervals. To the extent that these are the major drivers of patient behavior, we believe that the models can reasonably be pooled for analysis. As noted above, TIER.Net does not capture off-site medication collection visits for clients enrolled in DSD models, but only the facility-based visits that occur every 6 months, requiring a long follow-up window for retention to be documented. Since the TIER.Net system is not linked across facilities, moreover, “silent” or unrecorded transfers between facilities appear in the system as disengagement from care, likely underestimating overall retention in this study. We cannot know if this effect is differential or non-differential between DSD and comparison clients, however. In our analysis, assuming known transfers were retained in care at other facilities, then as many as 30% of those lost to follow-up in conventional care could have been silent transfers without changing the direction of results: we would still maintain a higher retention rate in the DSD arm. In addition, in South Africa, the CCMDD program, which provides the majority of medication for DSD models, requires a valid South African identification number, passport number, or asylum seeker number for registration. Anecdotally, a lack of a valid identification number for registration is often offered as a reason for not enrolling clients in the CCMDD program. In facilities with a large number of individuals without these numbers, such as those that serve immigrant communities, facilities may decide to pack medications themselves, at the facility, rather than relying on the CCMDD program. This could explain why some clients who appear in their medical records to be eligible for DSD are not enrolled. Similarly, the South African guidelines require clients to achieve control of all comorbid conditions before enrolling in a DSD model, but information on comorbid conditions is not captured in TIER.Net. For this reason, we were unable to incorporate this into the eligibility criteria for the inclusion of clients into the analysis. This would provide another explanation for clients who appear to be eligible but who are not enrolled.
The TTE methodology used in this analysis is a robust method for defining a comparator arm and determining outcomes in routine care data, but it cannot correct all issues that affect observational studies. It can help prevent issues like misaligned person time, poorly defined eligibility criteria, and unclear causal questions. It does not fully address the limitations of observational studies, however, as residual confounding can still be present. In our case, it is likely that many of those who were offered and accepted enrollment in DSD models were, on average, in better health, more motivated, or more empowered to remain in care than those who were not, suggesting that DSD participants may have had better outcomes than comparison participants even in the absence of DSD. This bias would overestimate the retention benefit of DSD over conventional care. The exclusion from DSD of otherwise-eligible clients without national identification numbers could also skew our results, either in favor of no difference (if clients without ID numbers also have good outcomes in conventional care, offsetting the bias mentioned above) or by overestimating the effect of DSD, if clients without ID numbers face greater barriers to remaining in care. Additional potential confounders include factors such as client choice, distance from the clinic, and history of adherence. Clients who choose DSD models may be different from those who do not, introducing a potential bias. Similarly, providers might be more inclined to enroll clients who live further away from the clinic, as they may perceive these individuals as having greater need for DSD models. Furthermore, providers may be hesitant to enroll clients with a known history of adherence issues, particularly those with a pattern of interrupted care.
Despite these study limitations, our findings, based on a robust observational methodology and large sample size, suggest that retention in care was slightly higher and viral suppression was similar among ART clients enrolled in DSD models in South Africa, compared to those in conventional care who meet typical DSD eligibility criteria. These findings should assure policy makers and program managers that the less intensive models of care that South Africa has introduced are unlikely to threaten the achievements of national HIV programs and that the non-clinical benefits they generate for the healthcare system, such as savings in provider time use, and for clients, such as less time and lower costs, are unlikely to be offset by poorer clinical outcomes.
Supporting information
S2 Table. Baseline demographics of trial-clients eligible for DSD models across all eight emulated target trials.
https://doi.org/10.1371/journal.pmed.1004489.s003
(DOCX)
S3 Table. Pooled adjusted risk difference (%) of retention in care and viral suppression by DSD enrollment at 12, 24, and 36 months.
https://doi.org/10.1371/journal.pmed.1004489.s004
(DOCX)
S4 Table. Pooled adjusted risk difference (%) of retention in care and viral suppression by DSD enrollment at 12, 24, and 36 months, adjusting for CD4 count at ART initiation.
https://doi.org/10.1371/journal.pmed.1004489.s005
(DOCX)
S5 Table. Age-stratified pooled risk differences for retention.
https://doi.org/10.1371/journal.pmed.1004489.s006
(DOCX)
S6 Table. Age-stratified pooled risk differences for viral suppression.
https://doi.org/10.1371/journal.pmed.1004489.s007
(DOCX)
S7 Table. Sex-stratified pooled risk differences for viral suppression.
https://doi.org/10.1371/journal.pmed.1004489.s008
(DOCX)
S8 Table. Sex-stratified pooled risk differences for viral suppression.
https://doi.org/10.1371/journal.pmed.1004489.s009
(DOCX)
S9 Table. Patient outcomes including reported TIER.
Net outcomes of those not retained within the trial outcome window period.
https://doi.org/10.1371/journal.pmed.1004489.s010
(DOCX)
References
- 1. Grimsrud A, Wilkinson L. Acceleration of differentiated service delivery for HIV treatment in sub-Saharan Africa during COVID-19. J Int AIDS Soc. 2021;24(6):e25704. pmid:34105884
- 2. Ehrenkranz P, Grimsrud A, Rabkin M. Differentiated service delivery: navigating the path to scale. Curr Opin HIV AIDS. 2019;14(1):60–5. pmid:30394947
- 3. Duncombe C, Rosenblum S, Hellmann N, Holmes C, Wilkinson L, Biot M, et al. Reframing HIV care: putting people at the centre of antiretroviral delivery. Trop Med Int Health. 2015;20(4):430–47. pmid:25583302
- 4. Ehrenkranz P, Grimsrud A, Holmes CB, Preko P, Rabkin M. Expanding the vision for differentiated service delivery: a call for more inclusive and truly patient-centered care for people living with HIV. J Acquir Immune Defic Syndr. 2021;86(2):147–52. pmid:33136818
- 5. Huber A, Pascoe S, Nichols B, Long L, Kuchukhidze S, Phiri B, et al. Differentiated service delivery models for HIV treatment in Malawi, South Africa, and Zambia: a landscape analysis. Glob Health Sci Pract. 2021;9(2):296–307. pmid:34234023
- 6.
HIV Coverage, Quality, and Impact Network (CQUIN). CQUIN 7th Annual Meeting Posters. CQUIN 7th Annual Meeting; November 13–7, 2023; Johannesburg, South Africa. Available from: https://cquin.icap.columbia.edu/event/cquin-7th-annual-meeting/posters/
- 7.
World Health Organization. Latest HIV estimates and updates on HIV policies uptake. Global HIV, Hepatitis and STI Programmes. 2020; 1–40. Available from: https://www.who.int/docs/default-source/hiv-hq/latest-hiv-estimates-and-updates-on-hiv-policies-uptake-november2020.pdf?sfvrsn=10a0043d_12
- 8.
South African National Department of Health. Adherence Guidelines for HIV, TB and NCDs. Pretoria, Republic of South Africa: Department of Health. 2023.
- 9. Liu L, Christie S, Munsamy M, Roberts P, Pillay M, Shenoi SV, et al. Title: expansion of a national differentiated service delivery model to support people living with HIV and other chronic conditions in South Africa: a descriptive analysis. BMC Health Serv Res. 2021;21(1):463. pmid:34001123
- 10. Dorward J, Msimango L, Gibbs A, Shozi H, Tonkin-Crine S, Hayward G, et al. Understanding how community antiretroviral delivery influences engagement in HIV care: a qualitative assessment of the Centralised Chronic Medication Dispensing and Distribution programme in South Africa. BMJ Open. 2020;10(5):e035412.
- 11. Luque-Fernandez MA, Van Cutsem G, Goemaere E, Hilderbrand K, Schomaker M, Mantangana N, et al. Effectiveness of patient adherence groups as a model of care for stable patients on antiretroviral therapy in Khayelitsha, Cape Town, South Africa. PLoS One. 2013;8(2):e56088. pmid:23418518
- 12. Wang M, Violette LR, Dorward J, Ngobese H, Sookrajh Y, Bulo E, et al. Delivery of community-based antiretroviral therapy to maintain viral suppression and retention in care in South Africa. J Acquir Immune Defic Syndr. 2023;93(2):126–33.
- 13. Lewis L, Sookrajh Y, van der Molen J, Khubone T, Maraj M, Sosibo P, et al. Long-term usage patterns and clinical outcomes in a community-based differentiated antiretroviral therapy delivery programme in South Africa. J Int AIDS Soc. 2023;26(7):e26141. pmid:40698878
- 14. Long L, Kuchukhidze S, Pascoe S, Nichols BE, Fox MP, Cele R, et al. Retention in care and viral suppression in differentiated service delivery models for HIV treatment delivery in sub-Saharan Africa: a rapid systematic review. J Int AIDS Soc. 2020;23(11):e25640. pmid:33247517
- 15. Fox MP, Pascoe S, Huber AN, Murphy J, Phokojoe M, Gorgens M, et al. Adherence clubs and decentralized medication delivery to support patient retention and sustained viral suppression in care: results from a cluster-randomized evaluation of differentiated ART delivery models in South Africa. PLoS Med. 2019;16(7):e1002874. pmid:31335865
- 16. Jo Y, Jamieson L, Phiri B, Grimsrud A, Mwansa M, Shakwelele H, et al. Attrition from HIV treatment after enrollment in a differentiated service delivery model: a cohort analysis of routine care in Zambia. PLoS One. 2023;18(3):e0280748. pmid:36917568
- 17. Uetela DAM, Augusto O, Hughes JP, Uetela OA, Gudo ES, Chicumbe SA, et al. Impact of differentiated service delivery models on 12-month retention in HIV treatment in Mozambique: an interrupted time-series analysis. Lancet HIV. 2023;10(10):e674–83. pmid:37802568
- 18.
The South African National Department of Health. Adherence guidelines for HIV, TB and NCDs: Policy and service guidelines for linkage to care, adherence to treatment and retention in care. Pretoria, Republic of South Africa: Department of Health. 2016.
- 19. Pascoe S, Huber A, Mokhele I, Lekodeba N, Ntjikelane V, Sande L, et al. The SENTINEL study of differentiated service delivery models for HIV treatment in Malawi, South Africa, and Zambia: research protocol for a prospective cohort study. BMC Health Serv Res. 2023;23(1):891. pmid:37612720
- 20. Osler M, Hilderbrand K, Hennessey C, Arendse J, Goemaere E, Ford N, et al. A three-tier framework for monitoring antiretroviral therapy in high HIV burden settings. J Int AIDS Soc. 2014;17(1):18908. pmid:24780511
- 21. Etoori D, Wringe A, Kabudula CW, Renju J, Rice B, Gomez-Olive FX, et al. Misreporting of patient outcomes in the south african national HIV treatment database: consequences for programme planning, monitoring, and evaluation. Front Public Health. 2020;8:100. pmid:32318534
- 22. Hernán MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Am J Epidemiol. 2016;183(8):758–64. pmid:26994063
- 23. Hernán MA, Sauer BC, Hernández-Díaz S, Platt R, Shrier I. Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. J Clin Epidemiol. 2016;79:70–5. pmid:27237061