Study design and setting: Thibela TB trial
This was a post-hoc analysis of data from the Thibela TB study, a cluster-randomised trial to investigate the impact of community-wide IPT on TB incidence and prevalence compared to standard of care among 15 clusters (mine shafts) of gold mine workers in South Africa [6]. Clusters were randomised to the intervention (TB screening offered to the entire cluster linked to treatment for TB disease or infection, as appropriate; eight clusters) or control (standard of care, seven clusters) arm [12]. The study methods are described in full elsewhere [12].
This analysis focused solely on the eight intervention clusters. Within each cluster, the same procedure was followed for enrolment and follow-up (Additional file 1: Figure S1). All mine workers within the cluster were invited to participate [13]. As 89 % of the workforce was estimated to be infected with M. tuberculosis, TST was deemed unnecessary [14]. Consenting miners were screened for TB using a symptom screen and a chest X-ray; a sputum smear and culture was ordered for miners with signs or symptoms suggestive of TB. Positive laboratory results and/or high clinical suspicion were reason for referral to the mine health services for further investigation and, if necessary, TB treatment. Participants without evidence of active TB or contra-indications to IPT were offered 9 months of IPT (300 mg isoniazid daily) [15]. During IPT, participants were seen monthly by study staff for dispensing of study medication, and a symptom screen for TB and adverse events. Symptomatic patients were referred to the mine health services for further work-up and management. After cessation of IPT, participants were evaluated for incident TB during 6-monthly or annual routine chest X-ray screening by occupational health services or after self-presentation with symptoms to the mine health services.
Population included in post-hoc analysis
All participants in the eight intervention clusters who were prescribed IPT at least once and who were permanent employees were included in this analysis. For analyses involving follow-up after the end of IPT, only participants who had completed the IPT period without a TB diagnosis, alive and still in the workforce, were included. In each cluster, follow-up was continued until 1 year after the end of the last participant’s intended IPT period.
Definitions and measurement of variables
Baseline characteristics of study participants were collected by a questionnaire at enrolment into the intervention and prior to initiation of IPT. At the request of the trade unions, no HIV testing was performed for study purposes. However, use of ART as concomitant medication was asked with respect to assessing possible adverse events. HIV prevalence among miners was estimated at 29 % in 2000 [16]. IPT dispensing dates were recorded; the number of monthly refill visits was used as a proxy of adherence (≥6 defined as optimal, 6 months being a recommended duration for IPT at the time the trial was designed [17]). The end of the intended IPT period was defined as 270 days from the first prescription date and the end of the actual IPT period as 30 days from the last dispensing date.
For estimation of the TB incidence rate, the start of the risk period was defined as the date of IPT initiation (objective 1); for the risk factor analyses the risk period started at the end of the intended or actual IPT period (objectives 2 and 3).
Miners with symptoms or signs suggesting TB were investigated and treated by the mine health services. Data on these episodes were obtained from record review. Routine diagnostic workup included sputum smear microscopy and a chest X-ray. Only one mining company routinely used mycobacterial culture to evaluate all miners with signs/symptoms suggestive of TB, the others only for those with a history of prior TB treatment. However, during the study, we endeavoured, following additional consent, to collect an additional sputum sample for smear microscopy and culture, though full coverage of this was not achieved. Cultures showing only non-tuberculous mycobacteria were not considered as TB. TB diagnoses were categorised as definite (2 positive smears or 1 positive culture for M. tuberculosis, or histological evidence of TB at autopsy in TB cases only ascertained post-mortem), probable (1 positive smear or culture with unidentified mycobacteria) or possible (clinical or radiological signs and symptoms, but no or negative smear/culture results). Medical records of participants with possible TB were reviewed by senior study clinicians, masked to study arm, to arbitrate on whether to include as incident TB.
Employment records were used to determine dates of employment and reasons for leaving the workforce (including deaths). Miners who died and underwent autopsy which revealed histological evidence of TB were included as incident TB cases.
The end of the risk period was defined as the earliest of TB treatment initiation (irrespective of the method of diagnosis), termination from the workforce, death or the end of follow-up. Participants who left the workforce due to being ‘medically boarded’ (discontinuing employment for a medical reason) were all evaluated by the mine healthcare system for active TB prior to the end of employment, and we therefore assumed that none of these represented missed TB diagnoses.
Statistical methods
The primary analyses were carried out using the intended IPT period to define the end of IPT (intention-to-treat (ITT), Fig. 1). Three sensitivity analyses were conducted: an as-treated (AT) analysis using the actual end date of IPT (Fig. 1), an AT analysis restricted to those with optimal adherence (defined as having been dispensed ≥6 months of IPT, see below), and a repeat analysis of both the ITT and AT analyses with a stricter TB case definition restricted to definite and probable TB diagnoses.
We calculated the overall TB incidence rate after initiation of IPT and period-specific rates during and after IPT. Poisson regression analyses adjusted for clustering by mine shaft were used to calculate TB incidence rates and to determine the rate ratio for TB after stopping IPT compared to during IPT. We investigated trends in TB incidence rates over time after the end of IPT using tests for linearity and departure from linearity.
Multivariable Cox regression analysis was used to investigate risk factors for TB following the end of IPT. We included factors we considered a priori as risk factors for TB (sex, age, previous TB, number of years in the workforce, type of work (reduced ventilation with underground work), housing (crowding in hostels), self-reported use of ART and number of times IPT was dispensed). The potential confounder (country of origin) was retained in the model if inclusion led to a change in hazard ratio estimate. In case of co-linearity, the more important of the co-linear variables was determined by literature review and retained. We considered previous TB might be on the causal pathway as it has been shown to be a risk factor for recurrent TB [18, 19]. We therefore examined the effect of dropping this variable in the final multivariable Cox model. The effect of excluding self-reported ART use was also examined. The final model was adjusted for calendar time.
Clustering was controlled for with a fixed effect for cluster in all regression analyses. Random effects analysis was not used due to small number of clusters [20]. Cluster estimates were not reported.
Departure from the proportional hazards assumption was tested in objective 3: we incorporated interaction terms between the time period of follow-up and the a priori risk factors. To maximise power, we used unadjusted analyses and two time periods: <12 and ≥12 months after stopping IPT. This time split was chosen to correspond with the expected timing of different underlying mechanisms (reactivation of inadequately treated latent infection <12 months versus reinfection ≥12 months), extrapolated from a study of recurrent disease after treatment [21].
For the AT analysis, these procedures were repeated using the actual end date of IPT as defined above. Two-sided P values in regression analyses were derived from likelihood ratio tests. Departure from linearity was tested for all ordered categorical variables. The analyses were conducted using STATA version SE 12.1 (College Station, Texas, USA).
Comparison between observed and modelled TB incidence in the first year after IPT
We applied base-case assumptions used in the published mathematical model of the Thibela study data to calculate a crude estimate of the average incidence rate of TB disease attributable to reinfection in the first year after IPT [22]. We did this overall and by cluster to allow for the varying annual risk of infection per cluster. We thereby assumed an annual risk of infection (averaged across clusters) before the intervention of 20 % per year and, based on outputs from the original model [22], a reduction in the annual risk of infection ranging between 11–20 % after IPT was introduced. For all clusters we assumed an HIV prevalence of 30 % of whom 25 % had a CD4 count below 200 cells/μL (and all of whom were eligible for and on ART), an initial pre-treatment loss to follow-up of 40 % and an average time to detection of 1 year. The age distribution and prevalence of silicosis depended on the cluster and were based on data collected as part of the study. The origin of these assumptions are explained in detail elsewhere [22]. Furthermore, also based on the original model, we assumed that reinfection, but no progression, could take place during IPT. We then calculated the observed TB incidence rate in the first 12 months after IPT (ITT), overall and per cluster, and compared these rates to the modelled estimates of TB incidence.
Ethical review
The Research Ethics Committees of University of KwaZulu-Natal, South Africa, and the London School of Hygiene & Tropical Medicine (LSHTM) as well as the South African Medicines Control Council and the South African Safety in Mines Research Advisory Committee gave approval for the Thibela TB study. The LSHTM ethics committee gave approval for this post-hoc secondary data analysis. Miners who participated in the intervention provided written or witnessed oral informed consent.