Patient retention in antiretroviral therapy programs up to three years on treatment in sub-Saharan Africa, 2007–2009: systematic review

Objectives To estimate the proportion of all-cause adult patient attrition from antiretroviral therapy (ART) programs in service delivery settings in sub-Saharan Africa through 36 months on treatment. Methods We identified cohorts within Ovid Medline, ISI Web of Knowledge, Cochrane Database of Systematic Reviews and four conference abstract archives. We summarized retention rates from studies describing observational cohorts from sub-Saharan Africa reporting on adult HIV 1- infected patients initiating first-line three-drug ART. We estimated all-cause attrition rates for 6, 12, 18, 24, or 36 months after ART initiation including patients who died or were lost to follow-up (as defined by the author), but excluding transferred patients. Results We analysed 33 sources describing 39 cohorts and 226 307 patients. Patients were more likely to be female (median 65%) and had a median age at initiation of 37 (range 34–40). Median starting CD4 count was 109 cells/mm3. Loss to follow-up was the most common cause of attrition (59%), followed by death (41%). Median attrition at 12, 24 and 36 months was 22.6% (range 7%–45%), 25% (range 11%–32%) and 29.5% (range 13%–36.1%) respectively. After pooling data in a random-effects meta-analysis, retention declined from 86.1% at 6 months to 80.2% at 12 months, 76.8% at 24 months and 72.3% at 36 months. Adjusting for variable follow-up time in a sensitivity analysis, 24 month retention was 70.0% (range: 66.7%–73.3%), while 36 month retention was 64.6% (range: 57.5%–72.1%). Conclusions Our findings document the difficulties in retaining patients in care for lifelong treatment, and the progress being made in raising overall retention rates.


Introduction
Although a great deal of research on daily adherence to antiretroviral therapy (ART) in sub-Saharan Africa has been published, long-term retention of patients in treatment programs has received comparatively less attention. ART has clearly been shown to be effective in reducing mortality among those who remain in treatment and adhere to therapy (Egger et al. 2002;Coetzee et al. 2004;Ivers et al. 2005;Laurent et al. 2005;Lawn et al. 2005), but under programmatic conditions, not all patients remain in treatment. In addition to known mortality while in care, some proportion of patients drop out of treatment programs and do not restart care elsewhere (Brinkhof et al. 2009;Fox et al. 2010). These patients are at high risk of morbidity and death within a short time. A valid measure for assessing the long-term success of ART programs should thus incorporate both mortality and loss to followup. In recent years, this measure has often been reported as the rate of retention, or the proportion of patients remaining alive and on ART at various time points after treatment initiation.
In 2007 we published the first systematic review of retention rates in cohorts in non-research settings in sub-Saharan Africa. In that review we demonstrated that, among over 74 000 patients representing 13 countries, the proportion of patients alive and on ART 2 years after initiation was approximately 60% (Rosen et al. 2007). The findings raised concern about high rates of attrition under programmatic conditions and suggested that more attention was needed to maintaining high rates of patient retention in both the months immediately following initiation when mortality is known to be high (Lawn et al. 2008), as well as over the whole course of patients' lives.
Since 2007, a number of publications have reported retention experience in cohorts not included in the earlier systematic review or reported on a longer duration of follow-up of cohorts previously included. In this paper, we present an updated systematic review of information published between mid-2007 and mid-2009. We reviewed recent data on retention and estimated the rate of all-cause adult patient attrition from ART programs implemented in service delivery settings in sub-Saharan Africa to determine whether rates of retention have changed since our initial estimates and to extend the previous estimates, which covered 24 months of follow-up, through 36 months.

Methods
To allow comparison between this review and the previous one, we maintained a similar methodology for this analysis as in our earlier review.

Search strategy
The current analysis included studies describing the results of observational cohort data from sub-Saharan Africa which reported on all adult HIV 1-infected patients who initiated first-line three drug combination ART, including those who discontinued treatment for any reason. We included cohorts from any facility, whether public, nongovernmental, or private, as long as they treated the general population using standard therapy. We excluded clinical trials and clinics serving specialized populations, such as workforces. To avoid doublecounting, we excluded reports of pooled data where it appeared that there was substantial overlap with patients also included in single-cohort reports. We required information beginning at ART initiation and a minimum, mean, or median follow-up of 6 full months (26 weeks). In some cases the average follow-up was not clearly specified but it could either be calculated from information provided or determined that it must be greater than 6 months. In cases where a cohort reported >50% retention at the last time point reported to, but no median duration of follow-up, the last time point reported to was used as the median. Reports had to include all-cause attrition rates for at least one of 6, 12, 18, 24, or 36 months after ART initiation. Studies that did not report on these time points but did provide enough information to calculate one of these rates were also included. When possible, children and non-naïve patients were excluded.
To identify studies for the current analysis, we searched Ovid Medline 2007-week 2, August 2009, ISI Web of Knowledge 2007-August 26, 2009 and Cochrane Database of Systematic Reviews 2nd quarter 2009) and four conference abstract archives (Conference on Retroviruses and Opportunistic Infections 2008International AIDS Conference 2007, HIV Implementers Meeting 2008 and International Association of Physicians in AIDS Care 3rd Conference on Treatment Adherence). The search in Medline, ISI, and the Cochrane Database combined the terms 'antiretroviral' and 'Africa' with any one of the following: retention ⁄ attrition ⁄ adherence ⁄ mortality ⁄ loss to follow-up ⁄ efficacy ⁄ evaluation or the term 'antiretroviral' and 'developing country' with either 'adherence' or 'mortality'. Conference abstracts were searched for any of the terms 'attrition' 'retention' or 'lost to follow-up' except for the IAPAC conference in which all abstracts were scanned.
SR identified the eligibility of all abstracts and journal articles that met our initial search terms and MF confirmed eligibility. For each study identified for inclusion we used a standard data extraction form to collect the relevant data. In cases where multiple reports described the same cohort, the one reporting to the longest time point or with the most complete information was used.

Definitions
We defined attrition from ART programs to include patients who died or were lost to follow-up. We defined retention to be the opposite of attrition (i.e. 1 -attrition). As reporting of patients who were still in care but had stopped taking ARVs was variable, we did not exclude these patients from the total retained in care. We accepted the varying definitions used in the reports for loss to follow-up and provide these definitions in Table 1. We excluded patients who were transferred to another facility from both the numerator and denominator of calculations of retention as we could not assess their outcomes.

Statistical methods
We first described each cohort and summarized its demographic and clinical characteristics by weighting the reported values by the sample size. We plotted the reported crude retention rates from each study at up to five time points: 6, 12, 18, 24 and 36 months. When retention rates were reported at time points different from those listed above or only for the median duration of follow-up, we ''-'' indicates information not reported in study. NGO, nongovernmental organization, ADOI, AIDS defining opportunistic infection, EPTB, extra-pulmonary tuberculosis. *Many public sector ART programs that did not report eligibility criteria likely followed national guidelines, which are typically consistent with WHO recommendations.
Median not reported, mean reported instead. àRequired at least one follow-up visit. §Includes non-naïve patients.
-Period during which study participants initiated ART; follow-up extends beyond dates shown. **Weighted the medians for those lost, died or retained by sample size. Sampled entire national program (adults). ààRepresentative sample of the national program. § §Includes 75 children.
--Rate from KM estimates only. Weighted average of pregnant and non-pregnant subjects. ***Study compared EFV and NVP; rates shown are for combined cohort. Includes non-naïve patients. Patients with missing records excluded. àààWeighted average of outcome groups. § § §Weighted averages of those on treatment and those not on treatment.
---Weighted averages of those on treatment at 12 months and those not.
****Total. applied the reported rates to the time point to which they were closest. For each study we calculated simple retention proportions (%RT) at each time point t as: where I 0 is all patients initiated on ART at the site; T t is all patients transferred out of care by time t; D t is all patients who died by time t; and LTFU t is all patients lost to follow-up by time t. This is the proportion of all patients initiated who did not transfer out of care who are still alive and in care at the end of the follow-up period. In cases where the first time point for which a study reported retention rates was later than 6 months after treatment initiation (25 studies representing 42 observations) we imputed the earlier missing time points. Because rates of attrition are unlikely to be constant over time (i.e. a linear decline), but rather likely to show a sharp decline over the first 6-months in care and level out sometime after 1 year, we first fit a linear randomeffects regression model using all the data available on current retention as a function of time, time squared, and time cubed with a random intercept for each study. The cubic form had a better fit than linear, quadratic, or half root forms. We then took the resulting predicted curve and calculated the proportion of the total 36month attrition that was expected to have occurred by each time period, with the constraint that retention could not increase over time. In this model, of total retention by 36 months, cumulatively 56% occurred by 6 months, 83% occurred by 12 months, 91% occurred by 18 months, 96% occurred by 24 months and 100% occurred by 36 months. This information was used to calculate the % retention at earlier missing time points as: where %RT i is the missing retention rate at month t i , 100% and %RT 1 are the two reported retention rates at baseline and time t 1 that bound %RT i (i.e. 0 < t i < t 1 ), and %TAT i is the % of total attrition up to the first time period reported (i.e. %RT 1 ) that occurred by time t i . For example, if the first retention proportion reported was 80% at 12 months, the percent of total attrition up to 12 months that occurred by 6 months would be: %TAT 6 ¼ ½0:56=ð0:83Þ ¼ 0:675 meaning that 67.5% of the total 20% attrition by 12 months occurred by 6-months. We would then estimate the 6-month retention proportion to be: For all time periods after the first reported retention proportion, if at least two non-consecutive time points of interest were reported but the study did not give retention proportions at one or more of the time points in between, we interpolated proportions (%RT i ) for the missing time points as a linear decline: where %RT i is the interpolated retention proportion at month t i and %RT 1 and %RT 2 are the two reported retention proportions at months t 1 and t 2 that bound %RT i (i.e. t 1 < t i < t 2 ). This was carried out for seven of the studies representing eight imputations.
As we previously demonstrated a relationship between duration of follow-up and reported retention rates (Rosen et al. 2007) (e.g. studies reporting only 6 months of followup on average reported lower retention at 6 months than the reported 6 month retention in studies reporting to 12 or 24 months), we first described retention rates in the current dataset by plotting them over time grouped by duration of reporting. We display retention rates at 6, 12, 24 and 36 months using forest plots to allow for visualization of the distribution of retention rates at each time point. We then summarized these rates using randomeffects meta-analysis (Egger et al. 2001) using a Freeman & Tukey (1950) arc-sin transformation of the retention proportions and standard errors and presented point estimates of retention and corresponding 95% confidence intervals.
To explore the impact of duration of reporting on overall retention, we projected the path of retention rates by extrapolating retention rates through 3 years (36 months) for all studies from their time of last reporting using three methods. We calculated a best-case scenario using last observation carried forward from the last time point reported through 3 years. We calculated a worst-case scenario by assuming that the attrition rates continued along the linear slope calculated between the first and last time point reported, with retention truncated at 0%. We calculated a midpoint scenario as the average of the two. In all scenarios, results were weighted by cohort size.
Finally, we looked for predictors of retention using linear regression. Potential predictors included median age in years (<36 vs. ‡36), % female (<60% vs. ‡60%), median baseline CD4 count (<100 vs. ‡100), duration of follow-up (£12 months vs. >12 months), patient payment required , and year of initiation of the cohort (<2004 vs. ‡2004). Because we had only 39 observations, all missing data were coded as unknown and included in regression analyses, but inferences were only drawn comparing the groups for which we had data. Two other potential, program-level predictors of retention are whether or not the treatment provider undertook active tracing of patients who were lost and made attempts to return them to care and how much pre-ART counseling was given to prepare patients for initiating and maintaining ART. Practices for preparing patients for ART and tracing those lost to follow up were not consistently reported, however, and therefore could not be included in the current analysis.

Results
Our search identified 632 papers and abstracts after removing duplicates. Of these, 491 were excluded on the basis of title or abstract, and an additional 108 were excluded after a full text screen and removal of overlapping cohorts and reports that contained insufficient information. Thirty-three sources remained that met our eligibility criteria (Figure 1), including 22 full text journal articles and 11 abstracts. These described outcomes for a total of 39 cohorts and 226 307 patients.
Several exclusions from the review and censoring of data should be noted. First, we omitted reports from the International epidemiological Databases to Evaluate AIDS (IeDEA) group that pools data from HIV ⁄ AIDS treatment programs. We excluded these reports because many of the cohorts included in IeDEA's analyses are described in more detail in the individual papers and abstracts in our review. Second, two studies, SA 1 and SA 2, reported outcomes to 48 months, and two studies, Botswana 2 and Cameroon 1, reported outcomes to 60 months. As these were the only cohorts reporting to time periods beyond 36 months, we censored data from these cohorts at the last time period reported to before 60 months (12 months for Cameroon 1 and 36 months for the remaining 3). And finally, we excluded several publications reporting data for cohorts from Malawi because the most recent report for Malawi, a conference abstract from July 2009 (Malawi 1), included data for the entire Malawian national ART program, thus encompassing the cohorts described in the other papers. Table 1 describes the characteristics of the cohorts. Nearly half the cohorts (49%, 19 ⁄ 39), representing 81% of the patients (183,486 ⁄ 226,307), were in southern Africa. The majority of treatment sites were either public sector (53%) or NGO (41%) facilities. About 35% of the 25 cohorts that provided information on payment reported that patients had to pay to receive care. The median years in which cohorts were enrolled for observation and in which observation ended were 2004 (range 1998-2005) and 2006 (range 2002-2007). Table 1 also reports median patient characteristics for each cohort. Patients were more likely to be female (median 65%, range 44%-100%) and had a median age at initiation of ART of 37 (range 34-40). All cohorts that reported starting CD4 count except one (Kenya 1) had a median below 200 cells ⁄ mm 3 (median 113, range 60-203). For the studies that reported it, the median of the reported follow-up times was 12 months.
The proportion of patients lost from each cohort at the end of that cohort's follow-up stratified by the cause of attrition, as well as the total proportion retained at each site, is shown in Table 2. After weighting for cohort size, loss to follow-up (LTFU) was the most common cause of attrition, followed by death (59% and 41% of total attrition respectively).
Total retention rates at each time point reported are presented in Table 3. For the 39 cohorts, attrition rates were reported at only one time point for 24 (61%); the median time point for these cohorts was 18 months (IQR 12-24 months). Total attrition at 12 months was quite variable, with a median of 22.6% and a range from 7 to   45% (Rwanda 1 and SA 10b respectively). There was little change in median attrition (27%) by 24 months, but the range narrowed slightly, from 11% to 35% in Rwanda 1 and Kenya 1 respectively. By 36 months median attrition increased to 29.6%, with estimates ranging from 13.0% to 36.1% (in SA 4b and Botswana 2 respectively).
To account for the variable times reported to, in all analyses presented below we interpolated retention rates at any time point where an estimate of retention was missing but a later time point was reported. Fifty estimates of attrition were interpolated, most at 6, 12, or 18 months (n = 25, 9, and 11, respectively). Figure 2a-d show the variation in retention rates at 6, 12, 24 and 36 months using forest plots. Using random effects meta-analysis to pool the data including the interpolated time points, we estimated the retention at  Figure 2 a-d Forest Plots of Reported Retention at 6, 12, 24 and 36 Months after Initiation of Antiretroviral Therapy in Sub-Saharan Africa*. *Pooled estimates were created using random-effects meta-analysis. Data include both actual reported rates for studies that reported to each time point and linear interpolation for studies which reported to a later time point but not the current time point. We were concerned that studies reporting only to shorter time points would have higher attrition at those time points than would studies that reported longer follow-up. Figure 3 shows retention rates stratified by last reported time point. Each time point shows variation in retention rates, but there is no clear picture of studies reporting at later time points having higher overall retention at earlier time points compared to studies reporting only to earlier time points. While the cohorts reporting only to 6 and 12 months show lower attrition with duration of time reported to, suggesting some bias may exist, the 8 cohorts reporting to 36 months show sharper declines in the first 6-12 months than cohorts reporting to 24 months.
To project the potential paths of retention over time, we conducted three analyses to extrapolate what would have happened to retention rates at later time points for cohorts reporting only to earlier time points (Figure 4). The first set of bars shows the best case scenario, in which the latest observation is carried forward with no further attrition, while the last set of bars shows a worst case scenario which assumes that retention continued at a linear rate. There was little variation in the estimates through 24 months as the retention midpoint between the two cases is 70.0% (best-case-worst-case range: 66.7%-73.3%). By 36 months, the retention midpoint between the two cases is 64.8% (best case-worst case range: 57.5%-72.1%).

Discussion
The global response to scaling up access to antiretroviral therapy in resource-limited settings has been rapid and dramatic and represents one of the largest public health successes in history. By the end of 2008, more than 4 million HIV infected patients had been initiated on life saving ART (Souteyrand et al. 2009). Even before the global economic crisis, however, the very large number of patients seeking treatment had begun to stretch resources and budgets. Now, as donors and governments face new pressure to reduce spending, difficult choices may have to be made about whether to use limited funds to initiate new patients on ART or to target funds towards keeping those already on ART alive and in care. Our findings document both the difficulties in retaining patients in care for lifelong treatment, but also the progress being made in raising overall retention rates. Our analysis of data from more than 22 500 patients initiated on ART in sub-Saharan Africa shows that in the period since our previous analysis, overall retention in ART programs 3 years after initiating patients on treatment averaged roughly 70% using a random effects meta-analysis of reported results. This finding, while not ideal, is notable in that most of the attrition occurred in the first 2 years on treatment, when both mortality (Lawn et al. 2008) and loss to follow-up are known to be high. After the first 24 months, attrition averaged about 5% per year.
The retention rate at 24 months we estimated in this analysis, 76% using meta-analysis, was higher than that estimated in our earlier review, which only analysed studies through 2 years. In that analysis we estimated 62% retention by 2 years. That analysis did not use a metaanalysis approach as was performed here but instead used a weighted average approach. This suggests that overall attrition by 2 years may be slowing as experience of scaling up ART accumulates and treatment programs mature and are better able to track patients. The greater attention to measuring and understanding loss to follow-up that is reflected in the large number of publications on this topic in recent years may be paying off as programs invest in interventions to track missing patients and return them to care.
We found substantial variation in retention rates reported over time but few predictors of overall retention that explain these differences and provide guidance on what determines attrition rates. While we were only able to estimate predictors of attrition in the first year after treatment initiation, we found that low median CD4 count and having fewer females as part of the cohort were predictive of higher attrition. Programmes with low median CD4 count would be expected to have more overall attrition because of mortality. As ART programmes expand, initiating patients with higher CD4 counts could lead to both improved outcomes on treatment as well as less overall attrition. Indeed, while our current analysis showed a very similar median initiating CD4 count to our previous analysis (128 in the current analysis vs. 132 in the previous analysis), others have found an increasing starting CD4 count over time (Keiser et al. 2008). Our analysis suggests that any such increases in baseline CD4 count may also be associated with increases in retention.
Loss to follow-up constituted the highest contributor to overall attrition, as was the case in our previous report. As the number of facilities offering ART expands, more of the patients reported as lost to follow-up may in fact have transferred informally to another facility. In our analysis, among studies which reported any transfers, about 7% of all patients who initiated care transferred to another facility. While most countries find it difficult to track patients from one facility to another, developing reliable referral systems that document transfers is essential to evaluating overall programmatic effectiveness.
For those patients who are lost but do not seek care elsewhere, mortality is expected to be high (Mocroft et al. 1997;Morgan et al. 2002;Badri et al. 2006). Brinkhof et al. (2009) conducted a systematic review of outcomes among patients lost from treatment programmes and estimated that about 40% (95% CI: 33%-48%) of LTFU patients had died, with much of the mortality occurring in the first 6 months after being lost to follow-up. Making efforts to get these patients back into care is important to the overall success of ART programmes, and developing ways to track and locate lost patients is essential to proper programme evaluation (Bisson et al. 2007). To further evaluate programmes, when available, method such as cross referencing with vital registration systems (Anglaret et al. 2004;Fairall et al. 2008) or adjusting mortality estimates statistically (Geng et al. 2008;Yiannoutsos et al. 2008;Fox et al. 2010) should allow for better estimates of programme impact as well as appropriate targeting of resources towards patient retention.
Our findings should be interpreted in the light of several limitations. First, we used data collected and reported by ART programs. Some misclassification of treatment outcomes likely existed, although we have little reason to believe these misclassifications would be anything more than random. Second, we integrated information from studies that reported at many different time points, and not all reported at a time point we were interested in. In these cases we applied the retention rate to the nearest time point, but this may have caused some bias in our overall estimates. Third, we were not able to determine whether the definition of LTFU used in specific studies influenced overall retention rates because of the variability in definitions, which made it impossible in many cases to determine how long patients were absent from the clinic before being classified as lost.
Fourth, systematic reviews can be subject to publication bias. Programmes managing cohorts with higher attrition rates might be less interested in publishing their results, and this would likely lead to an underestimation of overall attrition. At the same time, better-resourced programs may be both better able to retain patients in care and to conduct and publish research, particularly long-term analyses requiring consistent data collection and strong data analysis capacity. Both of these phenomena would lead to the bias that we found in our previous analysis, in which cohorts with higher attrition rates are more likely to publish outcomes only at earlier time points. We saw some evidence for this in the current analysis, as the average 6-month retention rates were lowest for studies that only reported to 6 months compared to the average 6-month retention rates from cohorts also reporting to later time periods. Our sensitivity analysis shows that if cohorts reporting only to earlier time points had continued along at the same rate of attrition as they reported at those earlier points, overall retention would be as low as 67% at 24 months and as low as 58% by the end of 36 months. While this sensitivity analysis may overestimate attrition, even our midpoint scenario (averaging the best-and worstcase scenarios) showed that overall attrition by 36 months is likely underestimated when using only published information.
Finally, our data were not extracted in duplicate as is recommended by the Cochrane Collaboration. We instead used single data extraction. This could have resulted in some studies being missed that potentially could have met the inclusion criteria.
In conclusion, we found that overall attrition by 24 months has likely decreased slightly since our earlier report, with overall retention by 24 months averaging 70%-77% and overall retention by 36 months averaging 65%-72%. Programmes initiating patients at lower CD4 counts also had higher rates of attrition than those initiating patients with higher CD4 counts. Active tracing of lost patients to return them to care and determine their vital status if not returned should be prioritized so as to keep overall retention high and to appropriately evaluate treatment programmes.