Overall 655 patients (27.2%) had at least one viral load measured between 6 months after starting ART and immunological failure. Viral load measurements were more common among patients who switched [159 patients (49.1%)] than among those who did not switch [496 patients (23.8%), P < 0.001]. Detectable viral loads (>400 copies/ml) were measured in 137 patients (86.2%) and 150 patients (30.2%), respectively (P < 0.001). The estimated hazard ratios for switching in patients who did and did not have viral load measurements were 0.13 (95% CI 0.02–0.93) and 0.40 (95% CI 0.13–1.31), respectively (P = 0.30 for difference), and the estimated hazard ratio including availability of a viral load test as a time-dependent confounder was 0.44 (95% CI 0.20–0.99). When only considering patients from sites with no viral load measurements in the MoH-CIDRZ programme the hazard ratio was 0.46 (95% CI 0.09–2.44). Finally, the causal effect of switching from the main analysis was insensitive to the different weight models, with hazard ratios ranging between 0.20 and 0.30.
The accuracy of the WHO CD4 cell count criteria to detect virological failure is poor . It is therefore not surprising that care givers are reluctant to switch patients to second-line therapy in ART programmes without access to viral load monitoring [5,14,15]. We aimed to clarify the causal effect of switching to second-line ART in such settings, using data from two large ART programmes in Zambia and Malawi. About 13% of patients with documented immunological failure switched to second-line ART during a median of 1.7 years of follow-up. We found that mortality from all causes was reduced by about three-quarters in patients who switched, compared to patients who did not, and that switching earlier reduced mortality compared to switching later. Our results also indicate that targeted viral load monitoring may further reduce mortality as it allows clinicians to selectively switch the patients who failed virologically and probably benefit most from second-line therapy.
The causal effect of switching to second-line therapy among immunologically failing patients should ideally be determined in a randomized controlled trial. However, it is unlikely that such a trial will ever be conducted: there must be ethical concerns about withholding an intervention that has a strong biological rationale, is standard in industrialized countries, and has been recommended by international bodies such as WHO. Observational data like those collected for this study may thus provide the best available evidence to inform clinical practice and public health decisions on whether and when to switch. Confounding by prognostic factors at baseline such as age, sex and clinical stage, and by time-varying factors such as CD4 cell count may, however, distort the results from observational studies [6,16]. We used inverse probability of treatment weighting to adjust for confounding by time-updated CD4 cell counts. This approach avoids the problem that CD4 cell counts are intermediate on the causal pathway from switching to death: confounding is controlled by weights rather than by inclusion of CD4 cell count as a covariate in the statistical model.
The reduction in mortality associated with switching was substantial, but it may have been diluted by unknowingly switching patients who met the immunological criteria for failure but had undetectable viral load. In a previous study from Malawi, viral load was measured in 203 patients with immunological or clinical failure: virological failure was present in 54% of patients, with high levels of viral resistance to nucleoside analog reverse transcriptase inhibitors and non-nucleoside analog reverse transciptase inhibitors [17,18]. Positive predictive values were even lower in a collaborative analysis of 10 treatment programmes from Africa and Latin America . In this study such dilution of the effect of switching will have been reduced by the targeted viral load measurement done in about half of patients who switched and about a quarter of patients who did not: viral load testing identifies virological failure and leads to switching in patients who benefit most. Indeed, in a sensitivity analysis we found that the reduction in mortality associated with switching was more pronounced in patients who had at least one viral load measurement compared to the patients without viral load measurements, although the difference failed to reach statistical significance. Whereas confounding by viral load testing remains a concern, analyses restricted to sites without viral load testing and analyses in which availability of a prior viral load test was treated as a time-dependent confounder both suggested that switching reduced mortality, although the former result failed to reach statistical significance. Our analyses were insufficiently powered to fully investigate these issues; we plan additional analyses as soon as more data on patients failing ART have become available.
Adherence is not consistently measured or reliably recorded in the participating cohorts and could therefore not be included in the analysis. Low levels of adherence are associated with poorer outcomes, both on first-line and second-line ART [15,19] and patients perceived by their providers to be nonadherent might also be less likely to be switched, resulting in unmeasured confounding. In our study, adherence will generally have been lower among patients in whom ART failed, compared to patients on nonfailing first-line ART. Interestingly, the proportion of men in our study was higher than in a recent analysis of patients starting first-line ART in the two programmes , in line with the notion that adherence is lower in men than in women [21,22]. Men were, however, as likely to be switched as women, indicating that documented or perceived adherence was not playing a major role in decisions to switch or not to switch. Opportunistic infections and cancers were also not consistently recorded. Like adherence, opportunistic events could influence decisions; however, such confounding would be expected to lead to underestimation the benefit of switching. Of note, in an ART programme in Mozambique, Malawi and Guinea-Conakry, only few patients met criteria for clinical failure without meeting immunological or virological failure criteria .
Less than 10% of patients were lost to follow-up after experiencing immunological failure. About twice as many patients are usually lost in the first year after starting ART  and about 40% of the patients lost to follow-up die in the following months . Although follow-up was limited after starting second-line therapy, it is likely that in the present study mortality among patients lost was also high: all patients had failed immunologically and their CD4 cell count was therefore low by definition. We controlled for informative censoring to the extent possible by incorporating in the weights a factor reflecting the inverse of a patient's estimated probability of remaining under follow-up at that time . Informative censoring remains a potential source of bias, however, particularly given the absence of death registries in Zambia and Malawi and the likelihood that some patients failed to return to the clinic because they had died .
We included two public-sector programmes involved in the scale-up of ART in Zambia and Malawi, with the Zambian programme contributing the vast majority (96.1%) of patients. The generalizability of our results may therefore be questioned. However, the two programmes follow national guidelines for the public health approach to ART, which is recommended by WHO , and are typical for many other programmes in the region. Results should thus be applicable to many patients in a region heavily affected by HIV. We nevertheless acknowledge that the two sites included in this study will not be representative for all programmes in the region. For example, they are equipped with electronic medical record systems, have access to regular CD4 cell counts and are involved in an international collaboration of HIV cohorts .
To our knowledge this is the first study that examined the causal effect of switching to a second-line regimen among patients with immunological failure in a low-income setting using marginal structural models. A previous analysis of data from USA  used marginal structural models to estimate the effect of switching to second-line therapy among patients with virological failure and found a hazard ratio of 0.81 in patients switching immediately compared to patients switching 3 months later. In a descriptive analysis of 11 ART programmes in sub-Saharan Africa, we found that mortality was higher in patients who switched compared to patients who remained on nonfailing first-line therapy, and substantially higher in patients who remained on immunologically or virologically failing first-line ART . These differences in mortality were not explained by differences in nadir CD4 cell count, age or sex, but time-dependent confounding was not considered .
In summary, our results suggest that in ART programmes in sub-Saharan Africa without access to routine viral load monitoring, patients meeting the WHO immunological failure criteria should be switched to a second-line regimen, and that delays in switching should be avoided. Our study also suggests that unnecessary switches might be avoided by targeted viral load; however, further research is needed to understand the barriers to switching patients who meet failure criteria, including, for example, the limited availability of second-line and third-line regimens, and the causal effect in patients with and without targeted viral load measurements. Finally, future studies should investigate at what CD4 cell count levels patients should ideally be switched, similar to the question of when ART should be started , and should examine long-term outcomes, including outcomes after second-line failure.
We thank all study participants and Marcel Zwahlen, Albert Mwango, Janne Estill, Gilles Wandeler and Lucy Campbell for helpful discussions and support.
Conflicts of interest
The study was supported by the National Institute of Allergy and Infectious Diseases (NIAID), Grant 5U01-AI069924–05 and a PROSPER fellowship to O.K. funded by the Swiss National Science Foundation (Grant 32333B_131629). The content is solely the responsibility of the authors and does not represent the official views of NIAID or the National Institutes of Health or the Swiss National Science Foundation.
Confounding by indication will bias results if switching is more likely in patients with lower CD4 cell counts at the time of immunological failure; patients who are also more likely to die. Similarly, CD4 cell counts measured after failure may both affect the probability of subsequent switching and mortality; unlike standard confounders, however, time-updated CD4 cell counts may themselves be affected by prior switching decisions. Such time-dependent confounding can be seen as a special case of confounding by indication , which cannot be controlled for by standard multivariable adjustment. These relationships can be visualized in a directed acyclic graph (DAG) , as shown in a simplified version in Figure A1.
To control for time-dependent confounding we estimated the causal effect of switching on mortality using two Cox proportional hazards marginal structural models of the counterfactual hazard of death under alternative switching times, making the following assumptions: no unmeasured confounding of the effect of switching on mortality, a probability greater than zero of switching and not switching at each time point, regardless of covariate values and correct specification of the treatment, censoring, lost to follow-up and outcome models.
The parameters of these marginal structural models were estimated using inverse probability of treatment weights (IPTW) . The inverse probability of treatment weights were stabilized and included additional censoring and loss to follow-up components . The treatment weight for a given time point was based on the inverse of a patient's estimated probability of having his or her observed switching history up to that time point, given his or her treatment and covariate history. Thus, for each quarter we estimated the probability of switching to second-line ART in that quarter among patients who had not already switched using Cox regression with predictors including all baseline covariates together with most recent CD4 cell count and the (lagged) CD4 cell count from the previous quarter. Most recent CD4 cell count refers to the last CD4 cell count before switching within a quarter.
For each weight model three different structures were considered: a simple linear model on the log-hazard scale including all predictors described above; a linear model on the log-hazard scale in which nonlinear associations were incorporated using natural cubic splines; and the best fitting linear model on the log-hazard scale based on AIC forward/backward stepwise selection procedure.
We used generalized additive models to assess linearity of the association between CD4 cell count and the log-hazard of switching.
P values correspond to tail probabilities of the coefficients’ posterior distributions.
Figure A1: Directed acyclic graph (DAG) illustrating time-dependent confounding in a setting with two time points and omitting baseline and unmeasured confounders. Arrows represent direct causal relationships between variables. CD4 cell count at a given time point (e.g. t = 0) not only determines if therapy is switched to a second-line regimen at some time point (e.g. t = 0 or t = 1, arrows a and b) but also influences CD4 cell count at a later time point (e.g. t = 1, arrow c) and mortality (arrow d). Switching at a given time point (e.g. t = 0) also influences CD4 cell count and switching at later time points (e.g. t = 1, arrows e and f) and mortality (arrow g). Similarly, CD4 cell count at time point t = 1 and switching at time point t = 1 influence mortality (arrows h and i). Finally, switching at time point t = 1 influences mortality (arrow j).
2. Keiser O, Orrell C, Egger M, Wood R, Brinkhof MW, Furrer H, et al. Public-health and individual approaches to antiretroviral therapy: township South Africa and Switzerland compared
. PLoS Med
4. Keiser O, MacPhail P, Boulle A, Wood R, Schechter M, Dabis F, et al. Accuracy of WHO CD4 cell count criteria for virological failure of antiretroviral therapy
. Trop Med Int Health
5. Keiser O, Tweya H, Boulle A, Braitstein P, Schecter M, Brinkhof MW, et al. Switching to second-line antiretroviral therapy in resource-limited settings: comparison of programmes with and without viral load monitoring
6. Sterne JA, Hernan MA, Ledergerber B, Tilling K, Weber R, Sendi P, et al. Long-term effectiveness of potent antiretroviral therapy in preventing AIDS and death: a prospective cohort study
7. Robins J, Hernan MA, Brumback B. Marginal structural models and causal inference in epidemiology
8. Hernan MA, Alonso A, Logan R, Grodstein F, Michels KB, Willett WC, et al. Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart disease
9. Cole SR, Hernan MA. Constructing inverse probability weights for marginal structural models
. Am J Epidemiol
10. Egger M, Ekouevi DK, Williams C, Lyamuya RE, Mukumbi H, Braitstein P, et al.Cohort profile: the international epidemiological databases to evaluate AIDS (IeDEA) in sub-Saharan Africa
. Int J Epidemiol
2011. [Epub ahead of print]
11. Goldman JD, Cantrell RA, Mulenga LB, Tambatamba BC, Reid SE, Levy JW, et al. Simple adherence assessments to predict virologic failure among HIV-infected adults with discordant immunologic and clinical responses to antiretroviral therapy
. AIDS Res Hum Retroviruses
12. Hernan MA, Wolfe F, Tilling K, Choi H, Sterne JA. Controlling for time-dependent confounding using marginal structural modelsA1: Fewell
. Z Stata Journal
13. Gelman A, Jakulin A, Pittau MG, Su YS. A weakly informative default prior distribution for logistic and other regression models
. Ann Appl Stat
14. Keiser O, Tweya H, Braitstein P, Dabis F, MacPhail P, Boulle A, et al. Mortality after failure of antiretroviral therapy in sub-Saharan Africa
. Trop Med Int Health
15. Pujades-Rodriguez M, Balkan S, Arnould L, Brinkhof MA, Calmy A. Treatment failure and mortality factors in patients receiving second-line HIV therapy in resource-limited countries
. J Am Med Assoc
16. Hernan MA, Brumback B, Robins J. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men
17. Hosseinipour MC, van Oosterhout JJ, Weigel R, Phiri S, Kamwendo D, Parkin N, et al. The public health approach to identify antiretroviral therapy failure: high-level nucleoside reverse transcriptase inhibitor resistance among Malawians failing first-line antiretroviral therapy
18. Hosseinipour MC, Kumwenda JJ, Weigel R, Brown LB, Mzinganjira D, Mhango B, et al. Second-line treatment in the Malawi antiretroviral programme: high early mortality, but good outcomes in survivors, despite extensive drug resistance at baseline
. HIV Med
19. Lima VD, Harrigan R, Bangsberg DR, Hogg RS, Gross R, Yip B, et al. The combined effect of modern highly active antiretroviral therapy regimens and adherence on mortality over time
. J Acquir Immune Defic Syndr
20. Keiser O, Chi BH, Gsponer T, Boulle A, Orrell C, Phiri S, et al.Outcomes of antiretroviral treatment in programmes with and without routine viral load monitoring in Southern Africa
21. Hawkins C, Chalamilla G, Okuma J, Spiegelman D, Hertzmark E, Aris E, et al.Sex differences in antiretroviral treatment outcomes among HIV-infected adults in an urban Tanzanian setting.AIDS
22. Bastard M, Koita Fall MB, Laniece I, Taverne B, Desclaux A, Ecochard R, et al.Revisiting long-term adherence to HAART in Senegal using latent class analysis
. J Acquir Immune Defic Syndr
2011. [Epub ahead of print].
23. Palombi L, Marazzi MC, Guidotti G, Germano P, Buonomo E, Scarcella P, et al.Incidence and predictors of death, retention, and switch to second-line regimens in antiretroviral-treated patients in sub-Saharan African sites with comprehensive monitoring availability
. Clin Infect Dis
24. Brinkhof MW, Dabis F, Myer L, Bangsberg DR, Boulle A, Nash D, et al. Early loss of HIV-infected patients on potent antiretroviral therapy programmes in lower-income countries
. Bull World Health Organ
25. Brinkhof MW, Pujades-Rodriguez M, Egger M. Mortality of patients lost to follow-up in antiretroviral treatment programmes in resource-limited settings: systematic review and meta-analysis
. PLoS One
26. Geng EH, Nash D, Kambugu A, Zhang Y, Braitstein P, Christopoulos KA, et al. Retention in care among HIV-infected patients in resource-limited settings: emerging insights and new directions
. Curr HIV/AIDS Rep
27. Gilks CF, Crowley S, Ekpini R, Gove S, Perriens J, Souteyrand Y, et al. The WHO public-health approach to antiretroviral treatment against HIV in resource-limited settings
28. Petersen ML, van der Laan MJ, Napravnik S, Eron JJ, Moore RD, Deeks SG. Long-term consequences of the delay between virologic failure of highly active antiretroviral therapy and regimen modification
29. When to initiate combined antiretroviral therapy to reduce mortality and AIDS-defining illness in HIV-infected persons in developed countries: an observational study. Ann Internal Med
30. VanderWeele TJ, Hernan MA, Robins JM. Causal directed acyclic graphs and the direction of unmeasured confounding bias
Keywords:© 2012 Lippincott Williams & Wilkins, Inc.
antiretroviral therapy; immunological failure; loss to follow-up; Malawi; marginal structural models; mortality; second-line therapy; treatment failure; viral load monitoring; Zambia