Secondary Logo

Journal Logo


Effect of antiretroviral treatment on the risk of tuberculosis during South Africa's programme expansion

Bachmann, Max O.a; Timmerman, Venessab; Fairall, Lara R.b

Author Information
doi: 10.1097/QAD.0000000000000806
  • Free



Tuberculosis (TB) is the commonest clinical manifestation of AIDS in Africa, where it aggravates TB epidemics by transmission of TB to HIV-uninfected as well as to HIV-infected populations [1]. Therefore, the priority and resources to be allocated to antiretroviral treatment (ART) expansion, and the stage of HIV/AIDS progression at which ART should be initiated, depend partly on the effectiveness of ART in preventing TB [2]. South Africa has the world's largest national ART programme, with government health services offering ART nationwide since 2004. By 2013, of an estimated 6.3 million infected with HIV, 2.6 million were receiving ART [3], mostly from government health services, and the incidence rate of TB notifications had started to decrease [4]. It is likely that ART expansion contributed substantially to the decline in TB incidence, but it may be difficult to quantify how much it did so, despite evidence from ART programmes [5–7] and epidemiological modelling studies [2,8,9].

Evidence of the effectiveness of ART in preventing TB is not directly available from placebo-controlled randomized trials, so observational studies are necessary to compare the incidence of TB between HIV-infected people receiving and not receiving ART. Such studies are however prone to confounding and selection bias, and therefore require good quality data on, and careful adjustment for, patient characteristics associated with both treatment and outcomes. Survival analysis of the effect of ART on time to TB should account for the competing risk of death. Conventional survival analyses, such as Cox proportional hazards regression, assume that the subsequent risk of TB in patients whose follow-up was censored due to death would have been the same as the subsequent risk of TB in patients whose follow-up was censored for other reasons. However, this assumption is probably false, because death in someone infected with HIV is an indicator that their disease was advanced, thus placing them at a higher risk of TB if they had not died. This can result in biased estimates of ART effectiveness. Competing risk regression avoids this potential bias [10].

A meta-analysis of the results of 11 studies in low and middle-income countries estimated that ART reduced the hazard or incidence of TB by 65%, and that ART effectiveness was not modified by patients’ pretreatment CD4+ cell count [5]. Although this is the best available evidence from developing countries, it has limitations: five of the 11 studies did not adjust for prognostic factors, including CD4+ cell count; the CD4+ cell count subgroup analysis was confined to four studies, and none of the studies accounted for the competing risk of death. The most rigorous observational evidence comes from 12 European and United States cohorts, analysed by the HIV Causal Collaboration, using marginal structural models to adjust optimally for time-varying covariates, including CD4+, so as to minimize selection bias [11]. That analysis estimated that ART reduced the hazard of TB by 56% overall, but was not effective in patients with CD4+ cell counts below 50 cells/μl. However, the latter results may not be applicable to low or middle-income countries such as South Africa that have much higher rates of exposure to TB infection.

We have worked with the Department of Health of the Free State province of South Africa since 2004, monitoring and evaluating the provincial HIV/AIDS programme. This provided exceptionally rich and comprehensive data on everyone enrolled in the programme, enhanced by linkage of data from medical records, laboratory results, the national mortality register and the province's electronic TB register. We have previously used these data to estimate that ART reduced the risk of TB by 39% during the first 20 months of ART roll-out [6]. However, that study did not account for the competing risk of death and did not link HIV and TB programme data, probably leading to under-ascertainment of incident TB. The aims of the present study were to estimate the effect of ART on the risk of the first TB episode after enrolment during the first 6 years of the Free State programme, and whether ART effectiveness differed according to CD4+ cell count or year of enrolment.

Materials and methods

Setting and patients

The study had a cohort design, comparing TB incidence during person-time without ART and after initiating ART. The study population was all HIV-infected people aged 16 years and over who registered with the Free State Provincial HIV programme since ART provision began in May 2004 until June 2010, who were followed until 2 August 2010. Of 97 476 enrolled patients, 23 401 were excluded because they were aged under 16 (7063), were HIV-negative (1839) or provided no longitudinal data because they had no clinic visits recorded after enrolment (8026), had no recorded national identification number for linkage with the national mortality register (4616), died on the day of enrolment (836) or had impossible dates recorded that could not be resolved, such as death before enrolment or TB diagnosis (1021).

The Free State Provincial HIV programme offered free ART to all eligible patients since May 2004. ART included either of two triple-drug regimens for most patients, with second-line drugs reserved for a minority with ART resistance. According to treatment protocols at the time, adults were eligible for ART when their CD4+ cell count was less than 200 cells/μl, they had had stage IV HIV infection (AIDS) or during the last 3 months of recruitment, were pregnant with a CD4+ cell count of less than 350 cells/μl. Patients not yet eligible for ART received routine care, such as regular CD4+ cell count testing, until they became eligible. Patients eligible for ART were referred to ART sites in hospital outpatient departments for initiation of treatment and review of ART prescriptions every 3–6 months. Resource constraints resulted in ART initiation often being delayed even while ART provision was being rapidly expanded [12]. The provincial TB programme followed WHO and South African national policies for diagnosis, treatment and registration [4].

The Research Ethics Committee of the University of Cape Town approved the study protocol. Although it was not feasible retrospectively to obtain consent for the use of medical records for research from the tens of thousands of participants, many of whom had died, ethical principles for the use of medical records for research without consent were followed [13]. That is, the research was a service evaluation of public health importance and did not influence individual patients’ care, and confidentiality of individuals’ identities and data security were strictly adhered to.

Data collection and linkage

The primary data source was the HIV programme's electronic medical record for each patient, including their characteristics at the time of enrolment such as name, address, date of birth, sex, previous or current TB and national identification number, and clinical information from each clinic visit. Clinic visit data included treatment, weight, CD4+ cell count and morbidity including TB. These data were recorded on paper forms by each patient's clinician, either a nurse or a doctor, and entered into a database by a clerk at the same clinic. Electronic clinic records were known to have incomplete data on CD4+ cell count, TB diagnosis and death; therefore, they were deterministically linked with the province's electronic laboratory CD4+ cell count reports and TB register and with the national mortality register. Linkage with laboratory data and the electronic TB register used name, date of birth, sex, national identification number and address. Linkage with the electronic TB register identified 9708 study patients with TB in addition to the 11 896 enrolled patients recorded as having TB in the HIV programme database. Linkage with the national mortality register, which is derived from death certificates and includes 90% of deaths countrywide [14], used national identification numbers that were recorded for 89% of patients on the Free State HIV programme database.

Following the national TB programme protocol, active TB disease was diagnosed by a positive sputum smear or, for extrapulmonary and smear negative TB, by a physician using all available diagnostic information. The date of diagnosis was defined as the start of the first episode after enrolment, recorded in the electronic TB register or, if not recorded there, as the first date that TB was recorded in the HIV programme's electronic medical records.

Patients’ prognostic characteristics used for statistical modelling were age at enrolment, sex, previous TB, CD4+ cell count and weight. Time-varying CD4+ cell counts and weights were extracted both for the time of enrolment and for the time at which ART started, together with the dates when these were measured (see Statistical analysis below). CD4+ cell counts were usually recorded approximately every 6 months. For CD4+ cell counts at enrolment and at the start of ART, respectively, the CD4+ cell count dated nearest to the enrolment or ART start date was used, but only if dated less than 6 months previously. Weights were usually recorded at each clinic visit and were extracted using the same rules as for CD4+ cell counts.

Statistical analysis

The prognostic characteristics at enrolment of patients who started ART and those who did not were compared using multiple logistic regression. For estimation of ART effectiveness, ART was modelled as a time-varying covariate, with follow-up time split into separate records during which patients had either initiated ART or not. Thirty-nine thousand, three hundred and twenty-six patients had two records each, one with and one without ART, 26 480 had only one record because they never started ART, and 4572 had only one record because they started ART on the date of enrolment. Because of uncertainty about medication adherence, patients who had started ART were assumed to continue receiving it, that is the estimated effect of ART was actually the effect of ART initiation.

In each record, follow-up was censored at the earliest of the first recorded TB diagnosis, death, the last clinic visit if a patient did not have a national identification number, 2 August 2010 (the day before the last date of linkage with the national mortality register), or, for follow-up without ART, the date of ART initiation.

CD4+ cell count data were missing for 15% (9661/65 806) of records without ART and for 18% (8022/43 898) of records with ART. Weight data were missing for 19% (12 406/65 806) of records without ART and for 25% (10819/43 898) of records with ART. These missing data were imputed using multiple imputation using chained equations [15] with Stata statistical software [16]. Imputation used the same explanatory variables to be used in the statistical models described below, including CD4+–ART interaction terms, and also the relevant outcome variables. Missing CD4+–ART interaction values were imputed using the improved passive approach described by White et al. [15]. For analyses of time to first TB episode, the relevant outcomes used for imputation were death, TB and cumulative hazard function for TB [15].

The effect of ART on hazard of TB was estimated using proportional hazards competing risks regression [5], with death as a competing risk, adjusted for age at enrolment, sex, previous TB, CD4+ cell count, weight, year of enrolment and district. ART–CD4+ cell count interaction terms were added to the model to investigate whether effects of ART varied with CD4+ cell count. Alternatively, ART–year interaction was added to the model to investigate whether ART effectiveness changed according to year of enrolment. ART, CD4+ cell counts and weights were time-varying covariates, recorded either at enrolment or start of ART (as described above). Multiple imputation of missing CD4+ cell counts and weights produced 10 data sets, and regression results from all data sets were combined using Rubin's rules [17]. Nonindependence of repeated observations on individuals was accounted for using Huber-White robust adjustment of errors [18]. Cumulative incidence functions for death and TB as competing risks were estimated and graphed separately for person-time with and without ART.

Secondary analyses were done to assess the robustness of the primary analyses: complete case analysis excluding records with missing CD4+ cell count or weight data; censorship of follow-up at 12 months; and Cox regression analysis of time to TB, not accounting for the competing risk of death. All analyses were carried out with Stata statistical software [16].


Of 74 074 participants in the study, 43 898 (62%) initiated ART and 30 176 (38%) did not initiate ART or initiated ART after being diagnosed with TB. Patients who were female, or had CD4+ cell counts below 200 cells/μl, or had heavier weights, or had TB previously, or enrolled in later years or from some districts were more likely to initiate ART (Table 1).

Table 1
Table 1:
Characteristics at enrolment of participants associated with antiretroviral therapy initiation.

Patients were followed for up to 6.5 years (median 1.3, interquartile range 0.36–3.2 years). During 78 202 person-years at risk with ART, 3858 first TB episodes occurred and, during 62 801 person-years without ART, 5669 first TB episodes occurred (incidence rates 4.9 and 9.0 per 100 person-years respectively, crude incidence rate ratio 0.55 [95% confidence interval (95% CI) 0.52–0.57)]. During the same periods, 5536 died after starting ART and 10 398 patients died before ART could be started [incidence rates 7.1 and 16.6 per 100 person-years, respectively, crude incidence rate ratio 0.43 (95% CI 0.41–0.44)]. Figure 1 shows the cumulative incidences of TB and of death, both of which were lower with ART than without ART. The cumulative incidence of TB or death after 70 months of follow-up was 57% without ART and 37% after starting ART.

Fig. 1
Fig. 1:
Stacked cumulative incidence of tuberculosis (unshaded) and death (shaded), with and without ART: competing risks regression model.

ART and CD4+ cell counts above 200 cells/μl were independently associated with lower risk of TB, and male sex and later enrolment with the programme were associated with higher risk, in a competing risks regression model (Table 2). The estimated effectiveness of ART within each CD4+ cell count subgroup is summarized in Table 3. For patients with CD4+ cell counts of 350 cells/μl or less, ART was associated with 28–56% reduction in risk of TB. There was no effect of ART for patients with CD4+ cell counts of 350 cells/μl or more, but the CIs for this subgroup were wide (Table 3). ART effectiveness increased with each subsequent year of enrolment, with no significant effect among patients enrolled in 2004 and 2005, and with ART associated with halving of the risk of TB among patients enrolled in 2010 (Table 3).

Table 2
Table 2:
Variables independently associated with the risk of first tuberculosis diagnosis after enrolment: competing risks regression modela.
Table 3
Table 3:
Effect of antiretroviral therapy on risk of tuberculosis within CD4+ and year of enrolment subgroups.

Removal of ART–CD4+ cell count and ART–year interaction terms from these models resulted in an overall subhazard ratio for ART of 0.67 (0.64–0.70) (Table 4). Compared with this, secondary analyses showed that complete case analyses overestimated ART effectiveness, as did Cox regression analyses that did not account for competing risk of death (Table 4). Censorship of follow-up at 1 year after enrolment also increased the estimated effectiveness of ART (Table 4), but did not change the estimated effects of ART within subgroups defined by year of enrolment.

Table 4
Table 4:
Sensitivity analyses: effect of antiretroviral therapy on risk of tuberculosis estimated with different models, with and without imputation of missing data.


The study shows that ART initiation was associated with a substantially reduced risk of TB in HIV-infected patients with CD4+ cell counts of 350 cells/μl or less, and that the effectiveness of ART increased steadily during the first 6 years of the programme. These results differ from previous cohort studies in high TB burden countries including South Africa [5], in which effectiveness estimates were greater and in which effectiveness was not modified by CD4+ cell count. These differences could be partly attributable to the practical problems of delivering ART through a large-scale public sector programme with overloaded primary care facilities, compared with smaller programmes often delivered by nongovernment or academic providers on a smaller scale. However, the differences in results are probably also attributable to the methodological limitations of previous research, which our study was intended to avoid.

The study shows big changes during the first 6 years of the programme, notably the increasing effectiveness of ART each year (Table 3). Patients who enrolled in later years were more likely to start ART during the study, despite having shorter follow-up after enrolment (Table 1), but their risk of TB also increased with time (Table 2). The increasing TB risk could be partly due to increasing exposure to other people with infectious TB disease, as South Africa's HIV and TB epidemics rapidly grew [4], and was worse in patients who had not yet started ART. It may also reflect increased detection of active TB disease as part of the clinical work-up to starting ART, and province-wide efforts to strengthen integrated case management for patients with HIV [19,20]. The increasing effectiveness estimates are unlikely to be a data artefact due to shorter follow-up of patients who enrolled in later years, because censoring follow-up at a maximum of a year did not change these year-specific results.

Major strengths of the present study, compared with previous research in high burden countries [5], are better TB ascertainment through linkage of provincial TB and HIV registers, better adjustment for prognostic factors enabled by linkage with electronic laboratory data and better ascertainment of death in those lost to follow-up by linkage with the national death register. The statistical methods were also an advance on previous studies. The competing risk regression analysis would have reduced bias in effectiveness estimates due to the risk of death being lower in patients on ART than in patients not on ART, which bias was probably inherent in the previous Cox survival analyses [5]. Multiple imputation of missing CD4+ cell count and weight data enabled all eligible patients to be included in analyses, further increasing the generalizability of the results. Our sensitivity analyses showed that conventional Cox regression and exclusion of patients with missing data overestimated ART effectiveness, compared to our primary analyses (Table 4). Finally, the exceptionally large sample size allowed effectiveness to be estimated precisely, with sufficient statistical power to identify factors modifying ART effectiveness. For example, with over 18 000 participants who started ART with CD4+ cell counts below 50, or between 50 and 99 cells/μl, the study had 93% power to detect a hazard ratio 0.95 between the two subgroups as statistically significant at the 5% level.

The study has several limitations. As an observational study of treatment effectiveness, rather than a randomized trial, it is inherently susceptible to confounding and selection bias. Statistical analysis with marginal structural models, as used by the HIV Causal Collaboration [11], could potentially have provided better control for confounding and selection bias, although it would not have explicitly accounted for the competing risk of death. In our study, data on repeated CD4+ cell counts were relatively infrequent before ART initiation, and viral load data were only recorded after ART initiation, limiting our ability to adjust further for these time-varying prognostic and selection factors. Treatment policy restricted ART initiation at CD4+ cell counts above 200 cells/μl to patients who had AIDS (or who were pregnant), so that patients in these CD4+ cell count subgroups who received ART were probably at a higher risk of TB than those who did not, leading to a biased underestimate of ART effectiveness in these subgroups. We found that, among participants who enrolled with CD4+ cell counts above 350 cells/μl, they were more likely to start ART if they were male, older or had had TB before, but more detailed clinical features of AIDS were not recorded electronically and could not be adjusted for. It was not possible to adjust for pregnancy and for sex together because only women can be pregnant. Some patients may have received isoniazid prophylaxis, which could have reduced their TB risk. Isoniazid was not routinely or commonly used in this programme, and was not recorded to allow further statistical adjustment. Another limitation, which the present study shares with other ART programme evaluations [5,11], is that it could only estimate the effect of ART initiation, and not the efficacy of ART with perfect adherence. The high mortality rate after ART initiation was probably partly caused by imperfect adherence or treatment. Finally, we did not have access to more recent data to assess ART effectiveness since 2010.

In summary, the study shows that ART was effective in preventing TB, although less effective than in previously studies, and that effectiveness increased with time. The study supports further expansion of ART as a foundation of TB control in South Africa and other high TB burden countries.


All authors contributed to the conception of the study, interpretation of results and writing the article, and read and approved the final manuscript. V.T. carried out the data linkage. M.O.B. and V.T. carried out the statistical analyses.

The authors thank the Free State Department of Health for provision of and permission to use these data for this research, and especially E. Kotze, health information director, and S. van der Merwe, TB programme director. We thank the two anonymous reviewers for their helpful suggestions.

Conflicts of interest

The authors declare that they have no conflict of interest. V.T. and L.F. were employed by the Knowledge Translation Unit, University of Cape Town, which received funding from the Free State Department of Health for monitoring and evaluation of the provincial HIV programme.


1. UNAIDS. 2013 UNAIDS report on the global AIDS epidemic. Geneva: UNAIDS; 2013.,85053,en.asp [Accessed 17 March 2015].
2. Pretorius C, Menzies NA, Chindelevitsh L, Cohoen T, Cori A, Eaton JW, et al. The potential effects of changing HIV treatment policy on tuberculosis outcomes in South Africa: results from three tuberculosis-HIV transmission models. AIDS 2014; 28 (Suppl 1):S25–S34.
3. Global Health Observatory Data Repository. Antiretroviral therapy coverage. World Health Organisation; 2015. [Accessed 17 March 2015].
4. WHO. Tuberculosis country profile: South Africa 2014. World Health Organization; 2014. [Accessed 17 March 2015].
5. Suthar AB, Lawn SD, del Amo J, Getahun H, Dye C, Sculier D, et al. Antiretroviral therapy for prevention of tuberculosis in adults with HIV: a systematic review and meta-analysis. PLoS Med 2012; 9:e1001270.
6. Fairall LR, Bachmann MO, Louwagie G, Janse van Vuuren C, Chikobvu P, Steyn D, et al. Effectiveness of antiretroviral treatment in a South African program: cohort study. Arch Intern Med 2008; 168:86–93.
7. Gupta A, Wood R, Kaplan R, Bekker L-G, Lawn SD. Tuberculosis incidence rates during 8 years of follow-up of an antiretroviral treatment cohort in South Africa: comparison with rates in the community. PLoS One 2012; 7:e34156.
8. Knight GM, Dodd PJ, Grant AD, Fielding KL, Churchyard GJ, White RG. Tuberculosis prevention in South Africa. PLoS One 2015; 10:e0122514.
9. Currie CSM, Williams BG, Cheng RC, Dye C. Tuberculosis epidemics driven by HIV: is prevention better than cure?. AIDS 2003; 17:2501–2508.
10. Fine JP, Gray RJ. A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc 1999; 94:496–509.
11. HIV-CAUSAL Collaboration. Impact of antiretroviral therapy on tuberculosis incidence among HIV-positive patients in high-income countries. Clin Infect Dis 2012; 54:1364–1372.
12. Ingle SM, May M, Uebel K, Timmerman V, Kotze E, Bachmann MO, et al. Outcomes in patients waiting for antiretroviral treatment in the Free State province, South Africa: prospective linkage study. AIDS 2010; 24:2717–2725.
13. Haines A, Ashcroft R, Coggon D, Coggon D, Coulter A, Doyal L, et al.. Personal information in medical research. Medical Research Council; 2000. [Accessed 17 March 2015].
14. Statistics South Africa. Mortality and causes of death in South Africa, 1997–2003: findings from death notification. Pretoria: Statistics South Africa; 2005.
15. White IR, Royston P, Wood AM. Multiple imputation using chained equations: issues and guidance for practice. Stat Med 2011; 30:377–399.
16. StataCorp. Stata: release 12. Statistical software. College Station, TX: StataCorp LP; 2011.
17. Rubin D. Multiple imputation for nonresponse in surveys. New York: Wiley; 1987.
18. StataCorp. Stata: release 14. Statistical software. Users guide. 2015; College Station, TX: StataCorp LP, 328–333.
19. Zwarenstein M, Fairall LR, Lombard C, Mayers P, Bheekie A, English RG, et al. Outreach education integrates HIV/AIDS/ART and tuberculosis care in South African primary care clinics: a pragmatic cluster randomized trial. BMJ 2011; 342:d2022.
20. Fairall L, Bachmann MO, Lombard C, Timmerman V, Uebel K, Zwarenstein M, et al. Effectiveness of task-shifting antiretroviral treatment from doctors to primary care nurses in South Africa (STRETCH): pragmatic cluster randomised trial. Lancet 2012; 380:889–898.

cohort studies; HAART; programme evaluation; treatment effectiveness; tuberculosis

Copyright © 2015 Wolters Kluwer Health, Inc.