Routinely used with ART in high-income countries, co-trimoxazole prophylaxis is usually discontinued when CD4 count exceeds 200 cells per μL because the primary goal is prevention of P jirovecii pneumonia.^12,13 Despite WHO guidelines, co-trimoxazole has been poorly used with ART in resource-limited settings; and data for its benefits (in addition to those of ART), toxic effects, and effect on ART adherence are scarce. In only one large retrospective cohort study in Malawi, investigators reported a 41% reduction in mortality during the first 6 months after ART initiation in clinics providing co-trimoxazole prophylaxis.¹⁴

Participants had variable exposure to co-trimoxazole; prophylaxis was never prescribed during follow-up for 324 (10%) participants, 1959 (62%) were taking it at ART initiation, and 896 (28%) started while on ART, with median 3·5 years (0·9–4·5) total use post-ART in those ever starting. Use at ART initiation varied with centre, year of randomisation, and initial combination treatment. Of 368 participants who died before end of follow-up, only 25 (7%) did not have complete co-trimoxazole history before death and were censored at their last clinic visit. 8128 (57%) person-years of follow-up were spent on co-trimoxazole. Use of prophylaxis differed substantially between the four centres (15%, 72%, 72%, 79%). Only 105 (<1%) person-years of follow-up were spent off ART. Reported ART adherence was high both in participants currently on and off co-trimoxazole, with no missed doses in the past 4 weeks reported at 5234 (83%) and 2426 (78%) visits, respectively, in the first 12 weeks, at 19 688 (93%) and 19 415 (87%) visits in weeks 12–72, and at 65 669 (93%) and 39 965 (91%) visits at more than 72 weeks of ART, excluding visits after randomisation to structured treatment interruptions.

Exclusion of follow-up after randomisation to structured treatment interruptions and upweighting of follow-up after continuous treatment gave 326 deaths (including eight upweighted) in 14 214 total person-years (2·3 per 100 person-years) included in analyses. 85 (26%) deaths occurred within 12 weeks of ART initiation (11·8 per 100 person-years). Present co-trimoxazole prophylaxis halved mortality in the first 12 weeks on ART, with no variation between participants on prophylaxis before screening (odds ratio [OR] 0·52, 95% CI 0·30–0·92, adjusted for baseline factors; p=0·02) versus those starting prophylaxis at the same time as ART (0·46, 0·25–0·84; p=0·01; heterogeneity p=0·69) or of variation between centres (heterogeneity p=0·35). Overall, using unweighted logistic regression and adjusting for baseline factors, but not controlling for time-dependent confounders, we showed that present co-trimoxazole use was associated with a 27% mortality reduction (0·73, 0·56–0·96; p=0·02; between-centre heterogeneity p=0·07).

As expected, mortality was higher in participants with low current CD4 cell count, haemoglobin concentration, or BMI, or with a WHO stage 3 or 4 event in the previous 4 weeks or any stage 3 or 4 event since randomisation (data not shown). These factors were also associated with increased probability of co-trimoxazole use, although extent of association varied between centres (data not shown). When we adjusted for these time-dependent predictors and randomisation to continuous treatment as factors within a regression model for mortality, the estimated OR for co-trimoxazole use was 0·62 (0·48–0·80; p=0·0002; between-centre heterogeneity p=0·50).

In a marginal structural model with inverse-probability treatment weights to control for time-dependent confounders, present prophylaxis reduced overall mortality risk by 35% (OR vs no present use 0·65, 0·50–0·85, p=0·001; table 2, figure). Benefit did not differ with randomised monitoring group (heterogeneity p=0·24). Furthermore, mortality risks did not differ between participants who had stopped prophylaxis and never users, and benefit did not vary with increasing time on co-trimoxazole in present users on ART (table 2, figure).

Benefit of present co-trimoxazole prophylaxis did, however, vary significantly with time on ART (heterogeneity p=0·04 in a flexible model for time on ART with cubic splines²⁰), falling from a 58% reduction in the first 4 weeks on ART to a 5% reduction in weeks 68–72 (every additional 4 weeks on ART to 72 weeks increased estimated OR by 1·05 [1·01–1·09]; p=0·03), with no effect subsequently (p=0·67). When we categorised time on ART to show this variation, mortality reduction was greatest in the first 12 weeks of treatment, sustained from 12–72 weeks, but not evident subsequently (heterogeneity p=0·02; figure). Adjustment for time on ART and time on co-trimoxazole simultaneously in present users showed that effect of time on ART was not confounded by duration of prophylaxis; similar mortality risk reductions were reported in the first 72 weeks on ART in participants on co-trimoxazole for fewer than 24 consecutive weeks (OR 0·50, 0·34–0·73) and more than 24 consecutive weeks (0·48, 0·28–0·83), and we noted no reduction in mortality after 72 weeks on ART in either group (0·91 [0·47–1·77], heterogeneity before vs after 72 weeks, p=0·12; OR 0·96 [0·63–1·47], heterogeneity before vs after 72 weeks, p=0·05 respectively). Estimated average 5-year survival in DART participants starting ART with CD4 counts of 15 or 150 cells per μL increased by 5% and 2%, respectively, with co-trimoxazole prophylaxis (figure).

We estimated effects of prophylaxis on mortality from primary, secondary, or tertiary causes regarded as potentially preventable by co-trimoxazole (92 deaths: septicaemia [39], pneumonia [16], severe brain syndrome as defined by the endpoint review committee [generally representing undiagnosed toxoplasmosis; 13], non-cryptococcal meningitis [8], diarrhoea [5], malaria [4], P jirovecii pneumonia [2], acute febrile event [2], toxoplasmosis [1], chronic pulmonary disease [1], or visceral abscess [1]), versus other causes (179 deaths: cryptococcus [31], tuberculosis [26], lymphoma or Kaposi's sarcoma [16], other [106]) versus unknown cause (55), separately. Present prophylaxis reduced risk of deaths from causes regarded as potentially preventable by co-trimoxazole by 21% (OR 0·79, 0·49–1·27), other deaths by 35% (0·65, 0·45–0·93), and unknown deaths by 48% (0·52, 0·27–1·00). In the first 72 weeks on ART, 66 (35%) deaths were from causes potentially preventable by co-trimoxazole compared with 92 (49%) due to other causes and 30 (16%) unknown; with similar mortality risk reductions associated with co-trimoxazole prophylaxis (OR 0·64 [0·37–1·10], 0·48 [0·30–0·77], and 0·29 [0·12–0·73] respectively). After 72 weeks, 26 (19%) deaths were from causes potentially preventable by co-trimoxazole, 87 (63%) were due to other causes, and 25 (18%) were of unknown cause; again, we found no difference in co-trimoxazole effect (OR 1·56 [0·59–4·14], 0·89 [0·53–1·50], and 0·93 [0·38–2·27], respectively).

All participants began ART with a CD4 count lower than 200 cells per μL; 2576 (62%) person-years of follow-up in the first 72 weeks and 2678 (27%) subsequently were spent with most recent CD4 count lower than 200 cells per μL. In the first 72 weeks on ART, we estimated similar mortality reductions in participants with current CD4 counts lower than 200 cells per μL and 200 cells per μL or more on co-trimoxazole prophylaxis (figure). After 72 weeks on ART we noted no benefit of present co-trimoxazole prophylaxis in participants with current CD4 counts lower than 200 cells per μL or 200 cells per μL or more (figure). Results were similar for deaths from causes regarded as potentially preventable by co-trimoxazole and for deaths from other causes. Alternative models including an additional CD4 count category of lower than 100 cells per μL (table 3), more than 350 cells per μL, or allowing for a non-linear effect of CD4 cell count also provided no evidence for variation in effect of co-trimoxazole prophylaxis with current CD4 cell count.

Oesophageal candidosis, cryptococcosis, and extrapulmonary tuberculosis were the most common WHO stage 4 events, contributing 159 (31%), 115 (22%), and 108 (21%) first events, respectively. Other diagnoses were individually less common. Only 22 (4%) stage 4 events were P jirovecii pneumonia (21 presumptive), of which 13 of 22 were on co-trimoxazole prophylaxis. We noted little evidence for a reduction in first new WHO stage 4 events after ART initiation with co-trimoxazole (table 2). The marginal benefit seen in past users is probably due to chance or incomplete adjustment for confounders in those stopping co-trimoxazole, and results of sensitivity analyses did not suggest that we had missed any benefit in present users. Results did not differ when oesophageal candidosis was excluded (data not shown). The effect of present prophylaxis on WHO stage 3 and 4 events, including recurrences of previous events as outcomes, was significant, but this small effect did not vary with time on ART (table 2). Heterogeneity existed between centres (p=0·02), and we were unable to obtain consistent estimates of the effect of co-trimoxazole on pneumonia alone (which accounted for 280 [22%] first WHO stage 3 or 4 events) across all four centres, suggesting variable diagnostic criteria or remaining unadjusted confounders.

In the first 12 weeks on ART, CD4 cell count increased by a median 73 cells per μL (32–128) and BMI by 0·76 kg/m² (0·00–1·60). Participants taking co-trimoxazole did not have greater CD4 cell count or BMI increases than did non-users in the first 12 weeks or subsequently (table 4); estimated CD4 increases were slightly reduced with co-trimoxazole, probably because of some residual confounding (difference vs non-users −3 cells per μL [−12 to 6]; p=0·50). We showed no effect of co-trimoxazole on haemoglobin or platelet counts, and observed a small significant reduction in neutrophil count in participants currently on co-trimoxazole after 72 weeks on ART (table 4). Only 22 (3%) of 650 serious adverse events during the trial were judged to be related to co-trimoxazole (ten definite or probable, 12 uncertain relation, all haematological, rash, or hypersensitivity).

Randomised trials are needed for the highest level of evidence,²² but in their absence patients still have to be managed according to the best available evidence from other studies, particularly those that are large, well done, and have clear, plausible, and consistent results. Since our study was observational, identification and appropriate adjustment for confounders is essential. Participants were in a randomised trial of management strategies, so the most important clinical and laboratory information used by clinicians for participant management was obtained systematically and prospectively, and we were able to do several sensitivity analyses using different models of inverse-probability treatment weights. In addition to the large cohort size, that we had four centres was invaluable for testing of these models since heterogeneity in effect of co-trimoxazole between centres suggested that confounders could remain. In models presented, results were consistent across centres and randomised monitoring strategies. Finally, censoring due to loss-to-follow-up was very low, and although we adjusted for censoring by using additional weighting, the effect was small. Since DART participants, who had advanced immunodeficiency and symptomatic disease, had similar characteristics to those of most patients starting ART in rollout programmes in Africa, our findings should be generalisable. Because our study used observational data we cannot guarantee that the results are free from bias, but with the systematic approaches used the possible direction and magnitude of any remaining bias are difficult to identify.

The survival benefit conferred by co-trimoxazole was restricted to present use, with no variation in benefit with use at or before enrolment or with increasing time on prophylaxis. The roughly 50% mortality risk reduction to 72 weeks after ART initiation is similar to that reported in ART-naive participants in randomised trials in resource-limited countries.^4,5,7,10 Beyond 72 weeks on ART, we observed no mortality benefits. One limitation of our data was that most present use beyond 72 weeks was in individuals on long-term co-trimoxazole, so we cannot distinguish between time from ART initiation and duration of effects of co-trimoxazole use after this point. Trials in ART-naive participants have typically had little follow-up beyond 72 weeks,^4,5,7 although Nunn¹⁰ reported a waning effect of co-trimoxazole prophylaxis with time, possibly related to falling adherence, in patients on treatment for concurrent tuberculosis. Stopping of co-trimoxazole prophylaxis after 72 weeks on ART might therefore be a reasonable strategy; randomised trials of different durations of prophylaxis would inform this issue. Although risk of malaria will remain (as for exposed adults not infected with HIV/AIDS), malaria in semi-immune adults is generally not serious and is non-fatal, and co-trimoxazole is not the prophylaxis of choice against malaria. In DART, absolute risk of death was increased at low CD4 cell counts, but we noted no variation in mortality risk reductions according to most recent CD4 cell count. Results of ART-naive studies have shown no evidence for heterogeneity in effect of co-trimoxazole prophylaxis by baseline CD4 cell count on severe morbidity⁵ or mortality.⁹

What co-trimoxazole prophylaxis is preventing, apart from malaria, is unclear from our data. In two trials of ART-naive adults, co-trimoxazole substantially reduced definitively diagnosed invasive bacterial disease with bacteraemia and malaria.^4,5 In children, lower respiratory tract infections were significantly reduced, but many died rapidly and most diagnoses were presumptive,^4,23 as seen in other non-randomised co-trimoxazole studies in African adults. In DART we obtained through structured summaries as much data as possible for deaths, which were then independently reviewed by an endpoint review committee. In 55 (17%) unobserved deaths, no cause could be assigned; in the remainder, primary, secondary, and tertiary causes were assigned with emphasis on confirmation of WHO 4 events because these were the primary endpoint. Although severe bacterial infections (as WHO stage 3 events) were systematically solicited and reported, they were not always microbiologically investigated. Thus, our data for cause of death cannot directly inform about mechanism of action, other than to note a substantial effect of co-trimoxazole prophylaxis on other causes, which could have been because of secondary sepsis.

Recurrent pneumonia diagnoses (individually reported as stage 3 events) have been reviewed by the endpoint review committee, because the revised WHO 2006 guidelines include two pneumonia episodes within 6 months as a stage 4 event.²⁴ However, insufficient events (n=13) were reported for a separate analysis. A key conclusion is that HIV studies in sub-Saharan Africa should document severe bacterial morbidity carefully, especially in patients with WHO stage 3 or 4 events. Although results of a recent study²⁵ suggested that almost all Mycobacterium tuberculosis isolates might retain in-vitro sensitivity to co-trimoxazole, we showed no evidence of an effect on pulmonary or extrapulmonary tuberculosis, similar to other WHO stage 3 or 4 events (data not shown).

Even if co-trimoxazole prophylaxis acts by partly or mainly reducing risk of mortality due to bacterial infections, why this mechanism should translate into mortality benefits before but not after 72 weeks is unclear. In particular, although one could postulate that severity of bacterial infections falls with increasing immune restoration, we did not show any variation in effect of co-trimoxazole by current CD4 cell count, and specifically no effect on deaths from causes regarded as potentially preventable by co-trimoxazole or other deaths in participants with low CD4 cell counts after 72 weeks. An alternative explanation, lent support by the absence of effect of prophylaxis on HIV-related disease progression and CD4 cell counts, is that although co-trimoxazole does not prevent HIV-related events, it might reduce mortality from them, irrespective of CD4 cell count. However, this explanation does not account for why the effect might occur before but not after 72 weeks from ART initiation. One hypothesis is that co-trimoxazole lowers bacterial load in the gut, thus reducing microbial translocation of bacterial lipopolysaccharides from the gastrointestinal tract into the bloodstream. Increased absorption of bacterial lipopolysaccharides raises immune activation,^26,27 and co-trimoxazole use at ART initiation could increase and accelerate reductions in immune activation, especially before ART has itself affected immune activation. In a few small studies from resource-limited settings, investigators have reported high pre-ART immune activation, which then decreases only gradually during the first 48 weeks of treatment.^28–30 Peripheral blood mononuclear cells were not stored in DART, so we cannot investigate this immune activation hypothesis further.