We developed a retrospective observational evaluation of adherence clubs. Accordingly, we built a Marginal Structural Model (MSM) using Inverse Probability of Treatment Weighting (IPTW) to estimate the intention-to-treat effect of adherence club participation. The analytical approach was developed specifically to address the obvious confounding bias whereby patients who were already doing well were more likely to be offered club participation. In the weighted analysis, club participation at any time following the start of the study was rendered independent of measured potential confounders. The study was nested within an ongoing cohort study of routine ART outcomes in Khayelitsha, Cape Town approved by the Human Research Ethics Committee of the University of Cape Town. [3] Individual patient consent was not needed, consistent with the South African Medical Research Council’s Guidelines on Ethics for Medical Research and the Declaration of Helsinki. Because this was a retrospective analysis of routine clinical service records, no additional data collection or procedures were undertaken from or on patients, all patient information was entered anonymously into the database using coded identification numbers, and no information that could reveal patient identity was entered into the database.

ART was first offered in Khayelitsha in 2001 at three community health centers serving an estimated population of 400,000. In 2011, over 20,000 adults and children had started ART in the sub-district. The clinical protocols have been closely aligned with international and national guidelines since inception. At the time that the adherence clubs were first introduced, the first-line regimen comprised stavudine, lamivudine and nevirapine or efavirenz, and patients received twice-yearly viral load and CD4 count monitoring. In a national guideline revision in April 2010 tenofovir replaced stavudine in the starting regimen, and viral load and CD4 count monitoring was reduced to annually beyond the first year on ART. [15] Historically, the majority of consultations have been with nurses, while doctors have been available to see very ill patients or receive referrals. Given the large patient load of several thousand patients on ART, patients do not consistently see the same practitioner from visit to visit. Facility-based lay counselors have provided structured patient preparation and adherence training at ART initiation and when adherence challenges have been identified. Home visits and patient support groups in addition to routine clinical care were a feature of the program in the first three years, but were not available to the majority of patients subsequently.

Adherence clubs were piloted in late 2007 at the largest of the community health centers (Ubuntu Clinic) to reduce the burden on health services by shifting consultations and medicine collections for stable patients to “clubs”. These groups are facilitated by non-clinical staff (counselors), who have previously been trained in facility-based patient preparation and support, and who receive additional training and mentoring. The meeting times are scheduled outside of busy clinic times, and consequently are more convenient for patients with much shorter waiting times. The resulting decongestion of the clinic allows nurses to spend more time initiating new patients on ART or with ill patients or those experiencing difficulties.

Participation is offered to clinically stable adult patients who have been on ART for at least 18 months. Guidelines further recommend that patients should have had a CD4 count of more than 200 cells/µl in the previous six months and have had sustained viral load suppression. Groups of 15 to 30 patients are formed and convene at the clinics during quiet times such as the early mornings or afternoons. Medicines are pre-packaged for each participant and brought to the group by a counselor who weighs the patients and administers a symptom-based general health assessment. Any patients reporting symptoms suggestive of illness, adverse drug effects or who have weight loss are referred back to the clinic to be assessed by a nurse. The counselor or experienced patients lead short group discussions on a range of health and other topics requested by the club participants. A nurse attends these groups annually to draw blood for viral load and CD4 count testing.

As adherence clubs was a pilot, only some of the stable patients were offered participation, based on the clinician’s enthusiasm for the model, and the opening of new clubs: 20 clubs were established during the pilot. The determination of which stable patients were enrolled into the clubs was therefore largely driven by service factors as opposed to patient factors. Although not formally documented, most patients with the option to transition to club-based care took up this option due to the prospect of much less time spent waiting for medicines.

Data were extracted from the electronic medical record system used in the HIV clinics, which was updated daily by data capturers from structured clinical records completed by clinicians or from club registers completed by the club facilitators. Rule-based consistency assessments were used to identify specific patients and data elements for review by quality assurance staff.

Following the inclusion criteria that clinicians were using to offer club participation, we restricted the study to adult patients (≥18 years old) who had been on ART for at least 18 months when the pilot started, or who reached 18 months on ART during the study period, and who’s most recent CD4 count was above 200 cells/µl. Patients entered the analysis at their first eligible visit after November 1^st, 2007 and exited at the date of outcome, date of censoring from follow-up or February 28^th, 2011.

The principal outcome was a combined outcome of time to either death or LTF. LTF was defined as not having any contact with the service in the six months following the analysis closure (between February 28^th and August 31^st 2011), and was determined to have happened at the date of the last contact with the service. Limited to patients who were virologically suppressed at study entry, the secondary outcome was the time to the first virologic rebound (>400 copies/mL).

The primary exposure/outcome relationships and variables considered a priori to be potential confounders are outlined in a causal diagram (Figure 1). These include age, gender, duration on ART and World Health Organization (WHO) clinical stage measured at study entry; CD4 count measured at ART initiation, study entry, at any given time (t), and t-6 months; and viral load suppression measured at study entry and at t.

We summarized patient characteristics as percentages for categorical variables and medians and interquartile ranges (IQR) for continuous variables. We computed crude rates for each outcome by patient characteristics at study entry, with 95% Poisson confidence intervals.

For each outcome, we used an weighted pooled logistic regression model to estimate the hazard ratio (HR) of club participation versus routine clinic-based care, including as covariates a time-varying indicator for history of club participation, month of follow-up (cubic splines with knots at the 5th, 25th, 50th, 75th and 95th percentiles) [16] and the baseline covariates (age in years, gender, CD4 at ART start and study entry per 100 cells/µl, Viral load suppression [<400 copies/ml] at study entry, duration on ART in months and WHO clinical stage). We used robust variance estimators.[17]–[19] The pooled logistic model closely approximates the Hazard Ratios because the probability of events occurring in any given month is small. [20], [21].

To estimate the inverse probability weights, we fitted a pooled logistic regression model for first participation in an adherence club. The model included month of follow-up, and the baseline and time-varying covariates (current CD4 count per 100 cells/µl, six month lagged CD4 count per 100 cells/µl and current viral load suppression <400 copies/ml). We used the predicted probabilities of club participation to compute the denominator of the IPTWs. The weights were then stabilized as previously described. [20] For the secondary outcome, the final stabilized weight included the probability of being uncensored by LTF.

Patients received a weight of one once they joined an adherence club, and were assumed to remain part of the club until the end of the study, akin to intention-to-treat analyses of randomized trials. The weighted analysis created a statistical pseudo-population in which the probability of entering an adherence club in each month was unrelated to the measured CD4 cell count or viral load, thus controlling for time-dependent confounders. [21] In order to show the magnitude of the measured time-varying confounding we compared the weighted hazard ratios for club participation to unweighted estimates for both outcomes, both adjusted for co-variates at study entry.

Finally, we explored the sensitivity of our estimates to alternative models specifications (categorizing continuous variables and including selected interaction terms) and truncation of the IPTWs. All analyses were performed with Stata v.12.

Improved retention in care might result due to the removal of these and other structural barriers to care, [25], [26] which are inherently improved with out-of-clinic models. Patients may also potentially take responsibility for tracing and linking group members back to care when they miss visits, and arguably be more successful at achieving this than facility staff. Finally the group dynamic itself may be an important contributor as was historically motivated. [14].

A previous descriptive evaluation of a patient-group-based model of care in Mozambique reported 97.5% of patients retained in self-forming patient groups, [7] which is similar to the proportion retained in the adherence clubs in Khayelitsha.

Virologic rebound was expectedly low in both groups due to the restriction to stable patients, but was lower in the club model. There is a strong association between virologic rebound and gaps or delays in medication collection. [27] The ease of medicine collection associated with club participation (patients collect pre-packaged medications during the club) and strong incentives to collect medicines at the time of scheduled club meetings could reduce potential gaps in treatment collection. In routine facility-based care the time spent at the facility sometimes precludes patients from keeping appointments due to other commitments.

Although the main motivation for the adherence club model was the need for more efficient care, the potential for the clubs to function as peer support groups is an important consideration. The value of support groups is widely argued, but the impact on virologic outcomes has not been systematically assessed. [28].

We have previously reported higher rates of loss to follow-up and virologic failure in adolescents and young adults in the study setting, [3], [29] consistent with the findings when adjusting our analyses by age, notwithstanding that the study was not explicitly designed to assess these associations. These findings and the specific challenges in supporting adolescents and young adults have also been noted elsewhere. [30].

The use of a combined outcome for LTF and death was necessitated due to the high proportion of unrecorded deaths that end up being defined as LTF, [3] but precluded the exploration of each process separately. Although the completeness of viral load results was similar between groups, the completeness was less than would be anticipated based on guidelines, possibly reducing the overall ascertainment of virologic rebound.

The majority of patients transitioning to club-based care were already responding well to therapy and were assumed to be adherent based on immunologic and virologic response. This indication bias when comparing to patients remaining in routine care was addressed in the analysis in two ways. First the analysis was adjusted for the same variables considered by clinicians when inviting patients to join a club, measured at the time that patients first became eligible for the clubs. Secondly, through the inverse probability weighting, any time varying confounding that might result from changes in these parameters after study entry was further adjusted for. Although simpler cohort analytical approaches showed similar benefits (as evidenced by the presented unweighted models), the approach based on IPTW’s was preferred due to the strong prior assumption that time-dependent confounding may be present. Assuming relevant associations with club participation have been correctly measured and included in the analysis, and there is no unmeasured confounding, the weighted results should approximate those from a randomized comparison. [20].

As with all observational studies, we cannot rule out insurmountable limitations, especially in adjusting for biases in who accessed adherence clubs. There remains the possibility that unmeasured patient-level factors such as openness to group-based care, and willingness and motivation to participate in the clubs could result in residual confounding. As described however, the determination of who entered clubs and who did not was largely the result of clinician practice and club availability rather than patient preference overwhelmingly patients offered the club model took up the option due to the perceived benefits. It remains a limitation of the study however that refusals and which club entries were patient-initiated were not recorded, and reasons for refusal where present could have introduced unmeasured confounding.

This analysis equates to an intention-to-treat study as after first entering a club, patients were considered to be in the club treatment group until the end of the study. This is necessitated as a return to clinic-based routine care is usually the result of clinical or adherence problems, which would increase the event rates in routine care if this follow-up time was apportioned to routine care, and bias the analysis in favour of the club model.

Finally, we were unable to formally assess potential risks associated with non-clinical care. While it is reassuring that retention and adherence were likely improved by the club participation, it is possible that some clinical diagnoses were missed.

Club availability has since been expanded and extended to all clinics in the sub-district, enabling ongoing increases in enrolment without additional clinical staff. Future operational research will need to verify that adherence strategies based on patient groups can be successfully adapted to the specific needs and characteristics of different service and community settings. The management of these groups is becoming increasingly challenging as the number of groups associated with a single clinic increases, requiring new management strategies and related research. Groups are now meeting in the community, introducing challenges in the delivery of pre-packaged drugs to community settings, and the collection, transfer and review of patient and program management data.

Patient adherence groups were found to be an effective model for improving retention and documented virologic suppression for stable adult patients in long term ART care in Khayelistha. Models based on patient adherence groups meeting outside of pressurized clinical consultation areas, and facilitated by non-clinical staff, are a promising approach to assist in the next wave of increased access to HIV treatment.