By the end of February 2007, approximately 18 months prior to study start, 113 health facilities, including 79 public facilities and 34 faith-based facilities, were providing ART, and according to the national monitoring and evaluation system, 9,693 adults were on ART: 3,628 for 6 months 3,086 for 12 months, and 2,979 for 18 months. From September 2008 to April 2009, we conducted a cross-sectional study in 20 of the 113 facilities.

Sample size calculations were based on the expected proportion of patients reporting perfect adherence 18 months after ART initiation as that proportion was expected to be lower than those at 6 and 12 months after ART initiation. Assuming an 18-month perfect adherence rate of 85%, a precision of ±5% or less, a design effect of 1.5, and a refusal rate of 25%, as well as application of a finite population correction factor and logistical considerations, a sample of 1,798 patients spread across 20 sites was considered adequate.

Study participation was restricted to adults aged ≥18 years at study enrolment who had initiated first-line ART at one of the study sites 6, 12 and 18 months [+/−2 months] prior to data collection and were still receiving treatment at their initiating site, or transferred into one of the study sites within 30 days of ART initiation and were still receiving it at that site. Patients who died, were lost to follow-up (defined according to national guidelines as no clinic visit for 90 days or more since the last documented visit), transferred to another clinic before study start were excluded, as were those who continued in care at their initiating site but had stopped ART prior to data collection.

Stratified multi-stage cluster sampling was used, with sites as clusters, and six strata according to type of facility (public and faith-based) and time on ART (6, 12 and 18 months). The first stage of sampling involved randomly selecting 14 public and 6 faith-based sites from the 113 sites providing ART services 18 months prior to study start, with the total number of each type of site determined based on the relative contribution of each of those types of sites to the total number of adults on ART in Rwanda recorded in the national monitoring and evaluation system. In the second stage of sampling, patient registers and charts at the selected sites were used to create sampling frames of all eligible patients per strata. The desired sample size per strata was determined based on the relative contribution of eligible patients in each stratum to the total number of eligible patients identified. The sample per strata was divided across sites based on the relative contribution of eligible patients at each site to the total number of patients required per strata. Potential participants were selected from the site-specific sampling frames using simple random sampling. Individual records were reviewed for each potential participant selected using random sampling to confirm that they met the study eligibility criteria. If a patient was found to be ineligible, she or he was replaced by another eligible patient from the appropriate site-specific sampling frame using a random replacement scheme.

Site staff otherwise unaffiliated with the study contacted the selected patients confirmed to meet the study eligibility criteria at home and invited them to return to the health facility to learn more about the study. For patients who presented to the clinic, study staff provided more details about the study, obtained written informed consent, and completed interviews. Every alternate participant was included in the viral load sub-sample, with blood draw occurring immediately after the interview. Participants received the equivalent of US $4.50 to cover transport to the clinic to participate in the study. Patients who did not present to the facility following the invitation by the site staff, refused to participate after presenting to the facility, or could not be located were not replaced.

Data were double data entered into a Questionnaire Design Software (QDS™) (NOVA Research Company, Bethesda, MD) database and analyzed using survey procedures in SAS Version 9.2 (SAS, Cary, NC) designed for complex survey data. Sampling weights accounting for the probability that a site would be selected from the 113 in the sampling frame (by type of site: public and faith-based) and the probability that a patient would be selected from the site-specific sampling frame of all patients on ART (by time on ART: 6, 12, and 18 months) were used together with finite population correction factors in all analyses to obtain nationally representative estimates. Descriptive statistics were used to describe study sites and patient characteristics by time since ART initiation. Principal components analysis considering information on dwelling conditions and household assets was used to create a poverty index which was then divided into tertiles, representing the poorest, middle and least poor respondents. An index of side effects was generated by summing scaled responses to whether each of 19 side effects were experienced and if so, their severity, in the 30 days prior to interview; the index was categorized based on the 25^th and 75^th percentile cut-offs, representing whether the respondent experienced no or few side effects, moderate side effects or severe side effects. If CD4 counts were not available from the visit at which ART was initiated, results from up to three months before or after the date of ART initiation were used.

Primary outcome measures were three-day adherence, 30-day adherence, treatment interruptions, and viral load at the time of interview. Adherence and viral load were treated both continuously and categorically using clinically relevant thresholds. Treatment interruptions were examined using rates per person-year on ART. The four measures were compared across time since ART initiation strata using descriptive statistics, as appropriate. Logistic regression analysis was used to identify patient- and site-level characteristics associated with two of those outcomes: sub-optimal 30-day adherence and detectable viral load. Sub-optimal adherence was defined as patient report of having taken less than 100% of prescribed doses in the 30 days prior to interview, and viral loads of more than 40 copies/mL were designated as detectable. Non-collinear factors significant at the α≤0.2 level in unadjusted models were introduced in multivariable models. Variables that were statistically significant at α≤0.05 and that contributed to the overall goodness of fit of the model were retained in the final models. If not independently associated with the outcome, time since ART initiation, age, sex and CD4 at ART initiation were forced into the final models regardless of their statistical significance. Characteristics associated with three-day adherence and treatment interruptions were not assessed due to insufficient variability in the data.

The study protocol was approved by the National AIDS Commission, the National Ethics Committee and the National Institute of Statistics in Rwanda, as well as by the Institutional Review Board at Columbia University. All participants provided written informed consent prior to interview, and when applicable, prior to blood draw for viral load assessment.

Of the 20 study sites, 14 (70%) were public sector and six (30%) were faith-based, as per the study design. Eleven (55%) were located in rural areas. There were 14 (70%) primary-level health centers and six (30%) secondary-level hospitals. Seven (35%) sites had started providing ART services in 2003 or 2004, seven (35%) in 2005 and six (30%) in 2006 or 2007. Six (30%) sites had active peer educator programs and ten (53%) conducted routine home visits for patients.

Table 1 shows the characteristics of the population on ART for 6, 12 and 18 months in Rwanda at the time of data collection. There were few statistically significant differences in the distribution of population characteristics by time since ART initiation. The majority of the population was female (65%) and, on average was 38 years old and had completed 5 years of school. Many patients did not have information in their medical records on their WHO stage (37%) or CD4 cell count (16%) at ART initiation, with more missing information on WHO stage observed among those who started ART more recently than those who had been on ART for a longer period of time (50%, 33%, and 24% missing for 6, 12, 18 month cohorts, respectively). Among those with information available on their disease stage at ART initiation, there was variation by time since ART initiation, with patients on ART for a longer period of time starting ART in a more compromised health stage (median CD4 counts of 258, 209, and 192 cells/µl for 6, 12, 18 month cohorts, respectively). The majority (89%) of patients were on Nevirapine-based regimens and on average took 1.6 pills per day. About one-quarter (26%) reported severe side effects in the 30 days before interview. Nearly (94%) all believed ART was “very effective”, more than half (57%) used at least one ART reminder tool, with alarm clocks used most often by 22% of the population, and 53% participated regularly in PLWHA support groups. Thirteen percent reported consuming alcohol on at least one day in the week prior to data collection.

As shown in Table 2, 94% of the population reported perfect three-day adherence. Perfect adherence in the 30 days preceding the interview was reported by 78% of the population. An additional 11% took 90–99% of all their pills in the 30 days prior to data collection, 7% took 80%–89, and 4% took less than 80%. The rate of treatment interruptions was one 1.1 per person-year on ART. Eighty-three percent of the population had an undetectable viral load, 9% had a viral load of 41–500 copies/mL and 8% had a viral load of >500 copies/mL. While none of these measures varied by time on ART, significant differences were observed by site (Figure 2), with perfect 3-day adherence ranging from 84% to 100% (p = 0.0013) across sites, perfect 30-day adherence ranging from 50% to 98% (p = 0.0598), treatment interruptions ranging from 0.1 to 3.1 per person-year on ART (p<0.0001), and viral suppression ranging from 70% to 100% (p = 0.0552).

Table 3 shows characteristics associated with reporting non-perfect 30-day adherence and having a detectable viral load in unadjusted and adjusted analyses. In adjusted models controlling for sex, characteristics independently associated with higher odds of non-perfect 30-day adherence were being on ART for 18 months vs. 6 months (AOR = 1.75, 95% CI [1.18–2.60]), being aged 18–24, 25–34 and 35–44 vs. ≥44 (AOR = 2.00, 95% CI [1.05–3.83]; AOR = 2.51, 95% CI [1.50–4.20]; AOR = 2.06, 95% CI [1.28–3.30], respectively), reporting severe vs. no or few side effects in the prior 30 days (AOR = 2.18, 95% CI [1.47–3.24]); having no documentation of CD4 cell count at ART initiation vs. having a CD4 cell count of <200 cells/µL (AOR = 1.65, 95% CI [1.03–2.64]); alcohol use (AOR = 1.55, 95% CI [1.01–2.38]); and attending sites which initiated ART services in 2003–2004 and 2005 vs. 2006–2007 (AOR = 1.82, 95% CI [1.07–3.09]; AOR = 2.11, 95% CI [1.23–3.62], respectively); sites with ≥600 vs. <600 patients currently on ART (AOR = 1.63, 95% CI [1.20–2.23]) or those with peer educators (AOR = 2.94, 95% CI [1.87–4.64]). Participation in an association for people living with HIV/AIDS (AOR = 0.60, 95% CI [0.42–0.87]); and receiving care at sites which regularly conduct home-visits (AOR = 0.60, 95% CI [0.42–0.87]) were independently associated a lower risk of non-adherence.

In adjusted models controlling for duration on ART, age and CD4 count at ART initiation, higher odds of having a detectable VL were observed among patients at sites with peer educators (AOR = 2.01, 95% CI [1.40–2.88[). Being female (AOR = 0.52, 95% CI [0.29–0.94]); participating in an association for people living with HIV/AIDS (AOR = 0.39, 95% CI [0.24–0.65]); and using a reminder tool (AOR = 0.45, 95% CI [0.29–0.70]) were negatively associated with having a viral load of more than 40 copies/mL.