The trial protocol was approved by the University of Cape Town Ethics and Research Committee and written consent was obtained from all participants.

Participants with first episodes of sputum smear positive pulmonary tuberculosis were enrolled in a randomized controlled trial of a micronutrient intervention (vitamin A and zinc), which has been reported elsewhere [17]. In brief, adults who attended two primary care tuberculosis clinics in Delft, Cape Town, South Africa between May 2005 and August 2008 were recruited. Rates of poverty and unemployment are high in Delft and the tuberculosis incidence rate increased from 496 to 737 cases per 100 000 during the study period. Study participants received standard anti-tubercular therapy for 5 days a week consisting of combination tablets contributing 600 mg rifampicin, 300 mg isoniazid, 1.6 g pyrazinamide and 1.1 g ethambutol for participants weighing 38 to 55 kg (Rifafour, Aventis Pharma Pty Ltd, Johannesburg). Doses were adjusted for participants weighing less than 38 kg or more than 55 kg. All participants were on directly observed therapy at the clinic or in the community. Research staff collected early morning sputum specimens weekly from participants for 8 weeks. The micronutrient intervention had no effect on . time to sputum smear or culture conversion up to 8 weeks on Kaplan Meier analysis (P = 0.15 and P = 0.38 respectively; log rank test) The current analysis is restricted to trial participants with a positive baseline sputum culture and with data on susceptibility to rifampicin and isoniazid. Participants with rifampicin resistance during the 8 week follow-up period, were excluded.

In keeping with national guidelines at the time, none of our HIV-infected participants received anti-retroviral therapy, but all received co-trimoxazole prophylaxis. All participants were screened for misuse of alcohol with the Cut down, Annoyance, Guilt and Eye-opener (CAGE) questionnaire [18]. Smoking status was categorized as either never or ever smoked. Body weight and height of participants was determined. The extent and size of lung cavities at baseline was assessed independently by two pulmonologists experienced in the use of the Chest Radiograph Reading and Recording System (CRRS)- and disagreements on radiographic reading were resolved by consensus [19].

Prior to anti-tubercular therapy, trial participants supplied one spontaneous sputum specimen and thereafter one early morning unassisted sputum specimen every week, up to 8 weeks. Sputum specimens were examined by means of fluorescent microscopy (Auramine stain) and graded according to international standards [20]. All specimens were decontaminated and cultured on liquid media using the BACTEC MGIT 960 system (Becton Dickinson, Sparts, Maryland). Positive cultures were confirmed as M. tuberculosis complex using an in-house PCR assay [21]. The date of culture conversion was taken as the date of the first negative culture provided there were no subsequent positive cultures. Participants whose first negative culture occurred at week 8, were regarded as converters. Routine phenotypic drug susceptibility testing for isoniazid and rifampicin of isolates was carried out using the MGIT 960 system from January 2008; isolates cultured prior to that date were tested retrospectively, but some isolates had lost viability.

During 2009, spoligotyping was performed on the majority of serial isolates collected during the trial period [22]. Strain-specific PCR assays were performed to detect M. tuberculosis infections by W-Beijing and non-W-Beijing isolates [23]. In order to optimize the baseline data, we captured the strain type for all the available baseline isolates, as well as those within the first 2 weeks of treatment, in order to categorize participants as either W-Beijing or non-W-Beijing strain types. Ten participants who had two strain types, one of which was W-Beijing, were categorized as having the W-Beijing strain for the analysis.

Data was captured in duplicate in Microsoft Office Excel 2003 and validated with SAS 9.2 software. To account for missing data, multiple imputation was conducted with the Amelia II software package [24],a reliable imputation procedure implemented in the statistical software package R,and five imputed data sets were generated. This number of data sets is generally considered adequate [25].

Cox proportional hazard regression modeling with the imputed datasets was used to estimate the hazard ratio of culture conversion for the following pre-determined baseline variables: age, gender, smoking status, alcohol abuse, sputum smear grading, sputum TTD, presence and extent of cavitation, M. tuberculosis genotype, HIV-infection, body mass index, haemoglobin, total lymphocyte count, serum C-Reactive Protein (CRP) (at baseline and change in concentrations within the first 2 weeks) and albumin. Hazard ratios are reported together with 95% confidence intervals based on Rubin's rules [25]. The proportional hazards assumption was verified by testing interaction effects of analysis time with baseline variables (α = 0.05) and graphically via log-log-plots [26].

To confirm the stability of our results we conducted a sensitivity analysis using Bayesian model averaging to each imputed data set (M = 20), and the results were combined by Rubins' rules [25], [27]. Model averaging is an alternative to model selection and combines estimates from different models, which contain different sets of variables, with the aim of achieving stable estimates that incorporate modelling uncertainty in addition to sampling uncertainty: Models that have a higher posterior probability to be correct receive a higher weight [26]. The combined weighted estimates can then be interpreted as a stabilized hazard ratio of culture conversion, which, together with adjusted standard errors, reflects the uncertainty in the modeling process. The posterior effect probabilities presented in our analysis state the probability that the hazard ratio in the Cox regression model for a variable is not one, thus providing a summary measure of how likely it is that a baseline variable has an effect and ranks the importance of the baseline variables.

A total of 154 participants with new smear-positive pulmonary tuberculosis were enrolled in the micronutrient trial, of whom 113 fulfilled the inclusion criteria for the current study (Fig. 1). Four of the 113 sputum isolates were mono-resistant to isoniazid. Complete data on all the variables investigated in the current analysis was available in 72 patients. The baseline characteristics of these participants before and after multiple imputation, is shown in Table 1. There were 11 HIV-infected participants (5 males, 6 females) and HIV-status was not determined in 9 who were lost to follow-up before the test could be done. All participants underwent chest radiograph examination at baseline, but 22 could not be located at the clinic site either due to mis-filing or loss (e.g. chest radiographs sent with participants who were referred to hospitals). Strain data was unavailable for seven participants whose sputum cultures were discarded before spoligotyping could be performed.

Participants had a high baseline bacillary burden, as reflected by the large proportion with a high sputum smear grading and a short median duration of TTD of 7 days (Table 1). Lung cavities were common and there was a high prevalence of smoking and alcohol abuse among participants. Spoligotyping showed that almost half of our participants had the W-Beijing strain type.

By week 6, nine percent of participants were lost-to follow-up and by week 8, this increased to 15%. Two cultures from participants who had not undergone culture conversion up to week 7, had contaminated cultures at week 8 and were categorized as not having converted. On Kaplan Meier analysis 50.4% of participants had undergone culture conversion by 8 weeks (data not shown).

Cox-proportional hazard regression analysis showed that TTD, the presence of lung cavities at diagnosis and ever smokers, were all associated with delayed culture conversion in univariate analysis. After adjustment for co-variates, TTD and lung cavities remained significant (Table 2). Although the W-Beijing genotype was not significant in univariate analysis, the confidence interval was suggestive of a delayed effect on conversion, even after adjustment for co-variates. INH mono-resistance was not associated with conversion in univariate analysis (HR = 0.59; 95%CI, 0.082; 4.28) and was not explored further due to the limited number of participants.

The estimated hazard ratios of the variables remained similar after model averaging (estimates not shown) confirming the stability of our main results. After Bayesian model averaging ever smokers, the presence of lung cavities and TTD had a posterior effect probability of greater than 80%, thus having a rather strong effect on sputum conversion, even after taking model uncertainty into account. A moderate effect of W-Beijing genotype was also present (Table 2).