Effectiveness of Antiretroviral Treatment in a South African Program:  A Cohort Study FREE

More than 5 million South Africans are infected with human immunodeficiency virus (HIV).¹ In sub-Saharan Africa, antiretroviral treatment coverage was estimated to be 23% among those with advanced infection in 2006.² Efforts to expand highly active triple antiretroviral treatment (HAART) provision in Africa may be slowed by concerns that HAART will be less effective because of inadequate health services and poor treatment adherence. It is unethical to estimate HAART effectiveness using randomized placebo-controlled trials, but cohort studies^3- 4 can potentially provide valid estimates. Studies^5- 10 in developing countries describe improved outcomes among patient cohorts receiving HAART. To our knowledge, no studies have compared patients receiving HAART with similar patients not receiving HAART, as performed in cohort studies^11- 12 from Europe and the United States. Descriptions of changes in outcomes among patients receiving HAART probably underestimate HAART effectiveness because, without treatment, health outcomes tend to deteriorate. Concurrent or historical comparisons of treated and untreated patients are likely to be biased because treatments tend to be selected according to patients' health status. However, with successive observations of changes in individuals, it is possible to adjust for these selection biases, as has recently been performed for European and United States cohorts.^9- 14

Following South African national policy,¹⁵ the Free State provincial health department established its Comprehensive Care, Management and Treatment of HIV and AIDS Program, which includes provision of HAART. This study estimates the effects of HAART on mortality, tuberculosis, CD4 cell counts, and body weight, in addition to the effects of cotrimoxazole prophylaxis on mortality in this program.

The study had a cohort design with repeated measures. The study population comprised all persons with HIV infection at least 16 years old and enrolled in the Comprehensive Care, Management and Treatment of HIV and AIDS Program from commencement (May 3, 2004) until December 31, 2005. Patients with at least 2 recorded program contacts were included in our study.

Health care was provided through 13 assessment clinics, 4 treatment sites, and 3 combined assessment and treatment sites that were located in all 5 districts of the province. After testing HIV positive at any Free State clinic, patients were referred to nurse-run assessment sites. Patients with World Health Organization (WHO) stage 4 AIDS or a CD4 cell count of 200 cells/μL or less were referred to a physician at a treatment site for possible initiation of HAART. For patients with active tuberculosis or other serious opportunistic infections, HAART was deferred until they were clinically stable or until the intensive phase of tuberculosis treatment was completed. Of patients receiving HAART, 68.3% started treatment with stavudine, lamivudine, and efavirenz, and 29.7% started treatment with stavudine, lamivudine, and nevirapine. Cotrimoxazole prophylaxis was indicated if the CD4 cell count was 200 cells/μL or less or if WHO stage 3 or 4 disease was present.¹⁶

Patients receiving HAART were monitored monthly at clinic visits and quarterly by a hospital physician. The CD4 cell counts were supposed to be measured at enrollment and every 6 months thereafter if receiving HAART or if the CD4 cell count was less than 500 cells/μL, or every 12 months if the CD4 cell count was at least 500 cells/μL. In practice, the median intervals between CD4 cell count measurements were 146 days (interquartile range [IQR], 72-228 days) among patients with initial CD4 cell counts of 200 cells/μL or less and 186 days (IQR, 117-226 days) among others. A total of 5162 patients had CD4 cell count measurements at baseline and subsequently at least once. A total of 8335 patients had weight measurements at baseline and subsequently at least once. The relative frequencies of CD4 cell count and weight measurements are shown in Figure 1 and Figure 2. Viral load and WHO stage were not recorded for untreated patients.

Patient information was written on paper forms by clinicians, entered into the province's computer system by trained data capturers, and downloaded to a central database every week. Data quality was monitored continually using 29 quality control routines.

We included patient data from the time of enrollment until December 31, 2005. Longitudinal data were arranged into monthly records. If there was more than 1 measurement per month, the mean monthly value was used. Patients were assumed to continue receiving HAART or cotrimoxazole therapy once started. Missing CD4 cell count and weight measurements, including baseline values, were imputed using multiple imputation by chained equations¹⁷ based on age, sex, months of HAART, months of follow-up, current tuberculosis status, baseline and current CD4 cell counts and weights, and baseline and current HAART and cotrimoxazole therapy. Mortality was tracked until December 31, 2005, by electronically linking program data with the national death register, using national identity numbers that were available for 84.1% of patients. The national death register is estimated to capture more than 90% of adult deaths in South Africa.¹⁸ We linked these data at 9 months and at 17 months after December 31, 2005. We also identified deaths recorded in medical records. The recorded deaths were as follows: 80.1% in the population register only, 4.2% in medical records only, and 15.7% in both. Ninety-eight percent of registered deaths were identified within 9 months of December 31, 2005, and 1.8% were identified afterward.

Tuberculosis diagnoses were made according to national guidelines.¹⁹ Pulmonary tuberculosis was primarily diagnosed by examining sputum smears. Only a physician could diagnose extrapulmonary or smear-negative pulmonary tuberculosis, taking into account other clinical or radiologic information.

We summarized patients' baseline characteristics, adjusting standard errors of variable estimates for clustering of estimates within clinics within districts.²⁰ The primary method of longitudinal data analysis was the use of marginal structural regression models to account for time-varying covariates and for intermediate variables and selection bias. The primary outcomes were death, newly diagnosed tuberculosis, and changes in CD4 cell count and weight. The primary explanatory variables for mortality were having started HAART or cotrimoxazole therapy, and the primary explanatory variable for tuberculosis was having started HAART. The primary explanatory variable for CD4 cell count and weight was the number of months since HAART was started. Baseline covariates were age, sex, district, clinic, initial weight, initial CD4 cell count, the presence of tuberculosis, previous antiretroviral treatment, and cotrimoxazole therapy at baseline. Time-varying covariates were latest CD4 cell count and weight, cotrimoxazole provision, and, for prediction of censoring, HAART. Analyses were performed at the person-month level, with multiple records per patient. Statistical analyses were performed using statistical software (STATA version 9; StataCorp LP, College Station, Texas).²⁰

Marginal structural models use probability weights to adjust for confounding by time-varying covariates that are also treatment outcomes.^11- 14 STATA codes were adapted from those by Fewell et al.²¹ Stabilized inverse probability of treatment weights was estimated using 2 logistic regression models. The first had HAART as outcome and baseline covariates and months of follow-up as explanatory variables. The second had HAART as outcome and had baseline, and time-varying covariates and months of follow-up as explanatory variables. Stabilized inverse probability of censoring weights was also estimated. The first model had censorship as outcome and had baseline covariates and months of follow-up as explanatory variables, while the second model had these plus time-varying covariates. These weights were combined to produce stabilized inverse probability of treatment and censoring weights. For models of cotrimoxazole effectiveness, equivalent methods were used to produce stabilized inverse probability of cotrimoxazole therapy and censoring weights. Once a patient started HAART or cotrimoxazole therapy, the probability of receiving treatment was assumed to be 1, so probability weights were unaffected by subsequent CD4 cell counts or weights.

The probability weights were then used in pooled logistic regression models with death or tuberculosis as the outcome and in linear regression models with current CD4 cell count or weight as the outcome. Analyses with death as the outcome were censored at death or at December 31, 2005. Analyses with tuberculosis as the outcome were censored at the time of diagnosis or the last visit. Analyses with CD4 cell count or weight as the outcome were censored at the last time they were recorded.

The logistic regression models with death or tuberculosis as the outcome were Cox proportional hazards regression models because subjects were removed from the population at risk after the outcome had occurred. Odds ratios from the weighted logistic regression models are hazard ratios in this context.^11- 12 Explanatory variables were baseline covariates, having started HAART or cotrimoxazole therapy, and months since the commencement of HAART (0 months if HAART had not been started). We used Huber-White robust estimation of standard errors for variable estimates to account for intrapatient clustering of outcomes over time.²⁰

In any regression model in which age, weight, CD4 cell count, or months of follow-up were explanatory variables, they were modeled as cubic splines with knots at the 5th, 25th, 50th, 75th, and 95th centiles of the respective variable. This was to optimize adjustment for confounding without assuming linear relationships between these variables and the respective outcome.

We performed 4 secondary analyses. The first was equivalent but unweighted pooled regression analyses for each outcome. The explanatory variables were baseline and time-varying covariates, months of HAART, or having started HAART or cotrimoxazole. Second, for a more homogeneous population, we repeated the analyses among patients with initial CD4 cell counts of 200 cells/μL or less. Third, we repeated the analyses after imputing missing CD4 cell counts and weights by carrying the latest measurement forward instead of using multiple imputation. This excluded patients with missing CD4 cell counts and weights at baseline (18.4% and 7.2%, respectively). For these analyses, we added 2 variables indicating whether the latest CD4 cell count or weight measurement was the baseline measurement carried forward. Fourth, we repeated the analyses of mortality excluding patients without national identity numbers.

The Free State provincial health department gave permission for the data to be analyzed for this study. Patients were not asked to consent for their routine clinical information to be captured. The study protocol was approved by the research ethics committees of the faculties of health sciences of the University of Cape Town and the University of the Free State. Data protection procedures were strictly followed, and patient identifiers were available only to health professionals or data managers working in the program.

By the end of 2005, 14 267 patients were enrolled and followed up for up to 20 months (median, 4 months; IQR, 1-9 months) after enrollment. Of these, 3619 (25.4%) received HAART for up to 19 months (median, 6 months; IQR, 3-9 months) after enrollment. The median number of contacts with a physician or nurse was 24 (IQR, 17-34) among patients who received HAART and 6 (IQR, 3-14) among all patients. A total of 6899 patients (48.4%) had initial CD4 cell counts of 200 cells/μL or less, and 3125 of these (45.3%) started HAART during the study period (Table 1); 594 others had initial CD4 cell counts of greater than 200 cells/μL and started treatment later. Among patients with initial CD4 cell counts of 200 cells/μL or less, those who received HAART were at enrollment more likely to be female, have received HAART previously, have current or previous tuberculosis, and have higher initial CD4 cell count, body weight, and a more advanced WHO stage. However, the association with sex was weak.

Two hundred eighty-eight deaths, 101 new tuberculosis diagnoses, and 1730 person-years were observed among patients after commencement of HAART. There were 2161 deaths, 863 new tuberculosis diagnoses, and 8981 person-years among patients not receiving HAART. Therefore, the crude incidence rate ratios associated with HAART were 0.61 (95% confidence interval [CI], 0.49-0.75) for tuberculosis and 0.69 (95% CI, 0.61-0.78) for death. Among patients with initial CD4 cell counts of 200 cells/μL or less, 267 deaths, 94 new tuberculosis diagnoses, and 1607 person-years were observed after commencement of HAART; 1340 deaths, 575 tuberculosis diagnoses, and 3811 person-years were observed among patients not receiving HAART. Therefore, the crude incidence rate ratios associated with HAART were 0.39 (95% CI, 0.31-0.48) for tuberculosis and 0.47 (95% CI, 0.41-0.54) for death. Of 4570 patients followed up for at least 1 year or until death, 2430 (53.2%) were known to have died, 2105 of these (86.6%) before they received HAART (crude odds ratio, 0.09 [95% CI, 0.07-0.10]).

After adjusting for other covariates, patients were more likely to start HAART if they had previously received HAART, had received cotrimoxazole therapy, had increased (baseline-adjusted) weight, or had an initial CD4 cell count of 200 cells/μL or less (Table 2). After adjusting for other covariates, patients were more likely to start cotrimoxazole therapy if they were male, were older, had started HAART, had lower CD4 cell counts, or had tuberculosis at baseline.

The dramatic effect of HAART on survival among patients with CD4 cell counts of 200 cells/μL or less is shown in Figure 3. The hazard of death was higher if patients were older, had lower CD4 cell counts, had decreased (baseline-adjusted) weight, or received cotrimoxazole therapy at baseline (Table 3). The hazard of death was lower if they had started HAART or cotrimoxazole therapy or had tuberculosis at baseline. In marginal structural models that included all patients, the hazard ratios for death were 0.14 among those receiving HAART compared with those not receiving HAART and 0.37 among those receiving cotrimoxazole therapy compared with those not receiving cotrimoxazole therapy (Table 4).

Crude trends in CD4 cell counts and weights for patients who did or did not receive HAART are compared in Figures 1 and 2. Marginal structural models that included all patients (with imputed missing covariate data) demonstrated that, for each extra month since the commencement of HAART, CD4 cell counts increased by 15 cells/μL and weight increased by 602 g (Table 4).

Secondary analyses provided similar effect estimates. However, the effects of cotrimoxazole therapy on mortality and of HAART on weight were smaller in unweighted models than in weighted models (Table 4). In the subgroup of patients with an initial CD4 cell count of 200 cells/μL or less, HAART had more effect on the presence of tuberculosis (hazard ratio, 0.33 [95% CI, 0.13-0.92]), the same effect on CD4 cell count increase per month (16 cells/μL [95% CI, 15-17 cells/μL]), and less effect on mortality (hazard ratio, 0.33 [95% CI, 0.17-0.64]) and on weight increase per month (407 g [95% CI, 281-583 g]). Results were insensitive to the method of imputation. When the latest observations were carried forward instead of using multiple imputation, the hazard ratios for death were 0.18 (95% CI, 0.14-0.24) for HAART and 0.38 (95% CI, 0.32-0.46) for cotrimoxazole therapy in weighted models that included all patients.

We tested the assumption that changes in CD4 cell count and weight were linear over time by adding quadratic and cubic time variables or by using cubic or linear splines of time, as well as by comparing R² values of alternative models. With CD4 cell count change models, linear time models fitted best, showing that CD4 cell counts increased steadily for up to 19 months of HAART, in keeping with the findings shown in Figure 1. However, weight change models that used linear splines of months of HAART had slightly higher R² values than linear models and showed that weight gain was greatest during the first 11 months of HAART, in keeping with the results shown in Figure 2.

Compared with patients with national identity numbers, the 16.6% without them were less likely to receive HAART (9.5% vs 28.4%) and cotrimoxazole therapy (36.3% vs 56.5%), were less likely to be diagnosed as having tuberculosis (4.6% vs 7.2%) or to die (11.2% vs 18.6%), and had higher median initial CD4 cell counts (187 cells/μL [IQR, 74-344 cells/μL] vs 159 cells/μL [IQR, 71-300 cells/μL]). However, survival analyses produced similar results if these patients were excluded, with hazard ratios for death of 0.14 (95% CI, 0.11-0.16) for HAART and 0.35 (95% CI, 0.31-0.40) for cotrimoxazole therapy in weighted models.

The study provides reassurance that HAART provided by South African public sector health services can be effective. The magnitudes of health improvements attributable to HAART are similar to estimates from dedicated nongovernment organizations and research-supported HAART services in South Africa^8- 9 and Haiti,⁶ countrywide treatment in Malawi,⁷ routine treatment in Europe and the United States,^11- 14 and indirect comparisons from randomized trials.^22- 23 However, it is alarming that more than half of the patients followed up for at least 1 year died and that 86.4% of those who died had not yet started HAART. These results demonstrate that expanding treatment access and avoiding delays in starting treatment are urgent priorities. The Free State provincial health department has made a commitment to expand access and to reduce waiting times, while striving to maintain the effectiveness of HAART.

This study extends the findings of other African cohort studies^3,5,7- 10 of HAART effectiveness because of the following strengths: (1) active follow-up of survival through data linkage with the national death register, (2) avoidance of selection bias by adjustment for time-varying CD4 cell counts and weights, (3) estimation of the effect of HAART on body weight, (4) estimation of the independent effect of cotrimoxazole therapy combined with HAART, and (5) inclusion of almost all patients enrolled in a provincewide government program.

The validity of the effectiveness estimates depends on valid and reliable measurement of all relevant confounders and outcomes, as well as correct model specification. A particular strength of this study was the successive CD4 cell count and weight measurements, which were biologically important prognostic and outcome variables and which influenced treatment selection. These measurements were made by routine services rather than under research conditions, but blood analyses were performed by a single national laboratory service with a single set of quality standards for its laboratories. A comparison of weighted and unweighted models (Table 3) suggests that unweighted models underestimated the effects of HAART on CD4 cell count and weight and the effect of cotrimoxazole therapy on death. The study was limited by the lack of viral loads and WHO staging for untreated patients, but these were not used to select treatment. We assume that 90% of deaths were captured through the national registry among the 84.1% of subjects with national identity numbers.¹⁸ Some deaths may have been missed among the 15.9% without national identity numbers. However, almost all deaths recorded in medical records were also registered centrally, and few deaths were registered longer than 9 months after the end of follow-up. If death registration was more complete for patients receiving HAART, then the effect of HAART on mortality would be underestimated.

We may have underestimated the incidence of tuberculosis, especially among patients not receiving HAART and seen less often. This is supported by the increased tuberculosis hazard associated with HAART in the crude analyses (Table 4). Therefore, we probably underestimated the effect of HAART on tuberculosis. It was surprising that tuberculosis at baseline was associated with a lower risk of death (Table 3). This was only true for patients who did not receive HAART. This could be due to selection bias because the sickest patients with tuberculosis may have died before enrolling in the program.

We also may have underestimated treatment efficacy by assuming that patients continued to receive HAART or cotrimoxazole therapy once started. By the end of follow-up, 3.1% of patients who had started HAART had not attended for at least 3 months, indicating that they had defaulted. Noncompliance would have biased the estimated effect on mortality more than the effects on other outcomes because mortality was tracked even if patients defaulted, whereas other analyses were censored at the last measurement or visit.

Our results reflect less than 2 years of follow-up among clinics and patients selected as being most ready to provide or receive HAART, so they cannot be directly extrapolated to larger-scale provision or to longer-term treatment. Our findings demonstrate that South African government health services and their patients are able to achieve good HAART outcomes but highlight the urgent need for increasing coverage and commencement of HAART.

Correspondence: Max O. Bachmann, PhD, Health Services Research, School of Medicine, Health Policy and Practice, University of East Anglia, Norwich NR4 7TJ, England (m.bachmann@uea.ac.uk).

Accepted for Publication: July 18, 2007.

Author Contributions: Dr Bachmann had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Fairall, Bachmann, Louwagie, van Vuuren, Bateman, Zwarenstein, and Chapman. Acquisition of data: Fairall, Bachmann, Steyn, Staniland, Timmerman, Msimanga, Seebregts, and Nhiwatiwa. Analysis and interpretation of data: Fairall, Bachmann, Louwagie, Chikobvu, Steyn, Staniland, Msimanga, Boulle, Bateman, and Zwarenstein. Drafting of the manuscript: Fairall and Zwarenstein. Critical revision of the manuscript for important intellectual content: Louwagie, van Vuuren, Chikobvu, Steyn, Staniland, Timmerman, Msimanga, Seebregts, Nhiwatiwa, Bateman, Zwarenstein, and Chapman. Statistical analysis: Fairall, Bachmann, and Boulle. Obtained funding: Fairall, Bachmann, Seebregts, Bateman, and Zwarenstein. Administrative, technical, and material support: Fairall, Bachmann, Louwagie, Steyn, Staniland, Timmerman, Msimanga, Seebregts, Bateman, and Chapman. Study supervision: Fairall, Bachmann, Steyn, Bateman, and Zwarenstein.

Financial Disclosure: None reported.

Funding/Support: This study was supported by research grants from ACACIA (Communities and Information Society in Africa), Connectivity Africa, and Governance, Equity, and Health Programs of the International Development Research Centre, Ottawa, Ontario.

Role of the Sponsors: The study sponsors had no role in the study design, data collection, data analysis, data interpretation, or writing of the report.

Additional Contributions: Krista Dong provided structured clinical records. Sonja Botha, MSc, and Clive Seebregts completed initial training with the data capturers and clinical staff. Gloria Rembe, BSc Hon, resolved many data queries. The informatics team in the Free State (especially Eduan Kotze, MSc, Susan Robertson, BSCSC Hons, Piet de Beer, BAdmin, Chantelle Macalagh, BCom, Thulani Mazibuko, Dip Management of Training, Bennie de Winaar, BSocSc, and Steve Cockeram) and the Antiretroviral Task Team in the Free State (especially Roeleen Booi, MBA, Mvula Tshabalala, MBChB, and Carol Mokobe) provided additional contributions. Debbie Bradshaw, PhD, provided death registration data. MediTech provided the software for data collection. The antiretroviral coordinators, physicians, nurses, data capturers, and administrative clerks completed forms, captured them electronically, assisted with the resolution of data queries, and provided valuable services to their patients, despite many challenges. The journal's referees provided many helpful suggestions.