DART trial design Methods: Patients, follow-up, data Use of cotrimoxazole prophylaxis Time-dependent predictors of use of cotrimoxazole prophylaxis The effect of cotrimoxazole prophylaxis on clinical outcomes DART (Development of AntiRetroviral Therapy) is a randomised trial of management strategies in symptomatic ART-naive adults with CD4<200 cells/mm3 initiating triple drug ART. 3316 participants have been randomised to – Laboratory and Clinical Monitoring (LCM): 12 weekly biochemistry and FBC, CD4 with all laboratory results returned to treating clinicians – Clinical Monitoring Only (CMO): 12 weekly biochemistry and FBC, with results only returned to treating clinicians if requesed for clinical reasons or a grade 4 toxicity; 12 weekly CD4, with no results returned to treating clinicians Following a pilot study in 137 patients, 813 patients who achieved CD4≥300 cells/mm3 after 48 or 72 weeks on ART entered a second randomisation comparing continuous therapy with structured treatment interruptions ((STI) 12 weeks on, 12 weeks off ART). The STI randomisation was terminated early on recommendation of DSMC in March 2006; all patients are now on continuous ART, or off ART for other reasons DART is running in 3 centres, 2 in Uganda (plus 1 satellite site), 1 in Zimbabwe Cotrimoxazole is prescribed at the discretion of the treating clinician Methods: Marginal Structural Models Regression models for use of cotrimoxazole prophylaxis Centre A B C D Total patients (n) (n=964) (n=942) (n=979) (n=294) On cotrimoxazole % person years 12% 77% 72% 71% Person years follow-up 2739 2714 2902 858 Table 2: Proportion of total follow-up time spent on cotrimoxazole prophylaxis by centre 3179 patients in DART were included (137 patients who took part in a pilot study of structured treatment interruptions of ART were excluded) 9214 years follow-up between January 2003 and March 2007 267 deaths; 84 (31%) within 12 weeks of ART initiation 369 first WHO 4 events and 1149 first diagnoses of malaria after entry Table 1: Characteristics of the included DART cohort at randomisation At ART initiation DART N=3179 Sex: female 2057 (65%) Age (years) (median, IQR) 36 (31-42) WHO stage: 2 3 4 644 (20%) 1794 (56%) 741 (23%) CD4 (cells/mm3) (median, IQR) 83 (29-137) Haemoglobin (g/dl) (median, IQR) 11.4 (10.3-12.7) Information on current medication (other than ART) including drug, start date and indication was collected at 12-weekly visits We distinguished between cotrimoxazole prophylaxis and cotrimoxazole treatment by duration of use (usually ≤14 days for treatment) and reason for use • Marginal structural models are causal models which can be applied to observational data which include time dependent confounders, which may themselves be affected by previous exposure – They are fitted in two stages: firstly each subject’s probability of having their own treatment history for each time period given their covariate history is estimated. Then the effect of treatment on outcome is estimated in a weighted regression model, with weights for a subject in a time period inversely proportional to probability of their observed treatment history • Data were split into 4-weekly intervals from DART randomisation. Cotrimoxazole prophylaxis use within any 4-week interval was defined as use for at least 7 days • Use of cotrimoxazole in the first 12 weeks after DART randomisation was assumed to depend only on baseline covariates • Logistic regression models were used to estimate the probability of cotrimoxazole in any 4week interval >12 weeks after randomisation adjusting for previous cotrimoxazole prophylaxis (in the last six 4-week intervals for intervals >24 weeks after randomisation; and in the last three 4-week intervals for intervals 12-24 weeks after randomisation), time since randomisation (as a fractional polynomial) and baseline covariates (age, sex, stage, CD4, haemoglobin at trial entry, recruitment year). Time dependent covariates considered were: current and lagged CD4, current and lagged haemoglobin, history of WHO stage 3/4 events since randomisation and entry into the structured treatment interruption randomisation – Models were fitted within each centre (denoted A/B/C/D) and trial arm (LCM/CMO) and separately for intervals >24 weeks after randomisation and 12-24 weeks after randomisation. Backward elimination using p<0.1 was used. We fitted all predictors identified in any of the models in a final model within centre and arm to compute inverse-probability treatment weights for each period • To adjust for censoring by loss to follow-up/end of follow-up the inverse probability of remaining uncensored was also estimated for each time period dependent on baseline and time-dependent covariates and cotrimoxazole history and included in the final weight • The effect of cotrimoxazole on outcome was estimated by weighted logistic regression model adjusting for baseline covariates and time since randomisation. Weights adjusted appropriately for time-dependent confounders. We combined estimates for ≤12 weeks and >12 weeks by including intervals ≤12 weeks with unit weight Use of cotrimoxazole prophylaxis by time since randomisation and calendar time Cotrimoxazole prophylaxis was often not continuous if started (patients stopped and started during follow-up); patterns of use differed by centre A B C D 12-24 weeks 24-48 weeks 48-72 weeks 72-96 weeks 96-120 weeks 120-144 weeks > 144 weeks % pe rs on -y ea rs o n CT X af te r 12 w ee ks Figure 2: Proportion of follow-up spent on cotrimoxazole prophylaxis by time since randomisation in patients who were not on prophylaxis at 12 weeks Time since randomisation Figure 3: Proportion of follow-up spent on cotrimoxazole prophylaxis by calendar time in patients who were not on prophylaxis at 12 weeks % pe rs on -y ea rs o n CT X