Estimating the contribution of transmission in primary healthcare clinics to community-wide TB disease incidence, and the impact of infection prevention and control interventions, in KwaZulu-Natal, South Africa

Background There is a high risk of Mycobacterium tuberculosis (Mtb) transmission in healthcare facilities in high burden settings. WHO guidelines on tuberculosis (TB) infection prevention and control (IPC) recommend a range of measures to reduce transmission in healthcare settings. These were evaluated primarily based on evidence for their effects on transmission to healthcare workers in hospitals. To estimate the overall impact of IPC interventions, it is necessary to also consider their impact on community-wide TB incidence and mortality. Methods We developed an individual-based model of Mtb transmission in households, primary healthcare (PHC) clinics, and all other congregate settings. The model was parameterised using data from a high HIV prevalence community in South Africa, including data on social contact by setting, by sex, age, and HIV/antiretroviral therapy status; and data on TB prevalence in clinic attendees and the general population. We estimated the proportion of disease in adults that resulted from transmission in PHC clinics, and the impact of a range of IPC interventions in clinics on community-wide TB. Results We estimate that 7.6% (plausible range 3.9%–13.9%) of non-multidrug resistant and multidrug resistant TB in adults resulted directly from transmission in PHC clinics in the community in 2019. The proportion is higher in HIV-positive people, at 9.3% (4.8%–16.8%), compared with 5.3% (2.7%–10.1%) in HIV-negative people. We estimate that IPC interventions could reduce incident TB cases in the community in 2021–2030 by 3.4%–8.0%, and deaths by 3.0%–7.2%. Conclusions A non-trivial proportion of TB results from transmission in clinics in the study community, particularly in HIV-positive people. Implementing IPC interventions could lead to moderate reductions in disease burden. We recommend that IPC measures in clinics should be implemented for their benefits to staff and patients, but also for their likely effects on TB incidence and mortality in the surrounding community.

Respondents were asked if they knew their HIV status. Respondents who reported being HIVpositive were asked if they were on anti-retroviral therapy (ART).
Respondent household size was extracted from existing DSA data.
Respondents were asked to list all indoor locations visited and transport used on an assigned day in the week before the survey. For each location visited (including their own home), they were asked for further details, including: • What type of location it was (options included 'own home' and 'clinic') • How long they spent there • How many people (adults and children) were there, halfway through the time they were there • How many of those people were children aged <15 years For each use of transport reported, they were asked for further details, including: • What type of transport it was • How long the journey took • How many people (adults and children) were on the vehicle at the start of the trip • How many of those people were children aged <15 years Respondents were also asked for additional details on their clinic visiting behaviour during the six months prior to the interview, including: • The number of days on which they had visited a clinic for their own health in the past six months • The number of days on which they had visited a clinic for on the behalf of someone else (e.g. to collect a prescription) in the past six months, not included any visits that were also made for their own health • The number of days on which they had accompanied someone else to a clinic in the past six months, not including any visits that were also made for their own health and/or on behalf of someone else Finally, respondents were asked when their last visit to a clinic was, and, if it was within the past two years, they were asked for the following information about their last visit: • How long they spent at the clinic • How many people (adults and children) were there, halfway through the time they were there • How many of those people were children aged < 15 Further details of the social contact survey are given in McCreesh et al 1 .

Analysis
For each location visited on the assigned day, adult contact times were calculated as follows. Firstly, the number of adults present was calculated as the reported total number of people present, minus the reported number of children present. If this gave a value less than zero, it was set to missing. The number of adults present was then capped at 100, as above this value, it is unlikely that the respondent had sufficient contact with each adult present to allow transmission. The capped number of adults present was then multiplied by the duration of time that the respondent reported spending in the location, to give the adult contact time.
Estimates generated using the data on the respondent's last clinic visit were weighted by the reported number of clinic visits in the past six months.
Respondents who reported being HIV-positive were considered to be HIV-positive. Otherwise, respondents were considered to be HIV-negative/unknown.

Recruitment
Of the 3090 people sampled for UO, 1723 (56%) were successfully contacted, 298 (10%) were dead or reported to have out-migrated, 1071 (35%) could not be contacted. Of those successfully contacted, 1704 (99%) completed an interview (Table S1).  Table S3 shows the estimated mean annual number of visits made to clinics, by sex, age, and HIV status, estimated from data on reported clinic visits in the past day, and in the past six months.

Sampled (%) Contacted
Overall, there is little difference between the estimates calculated using the data collected using the two different recall durations. The exception to this is the estimates by sex, where there is a large difference in mean annual clinic visits by sex using the six-month recall data, but not the one-day recall data. However, the confidence intervals for the one-day recall estimates contain the estimated values for the six-month recall.
As there is no evidence that recall bias has had a large effect on the estimates, the six-month recall data are used to parameterise clinic visiting rates in the model, due to their greater precision.

Contact in other locations
Other locations are defined as indoor locations other than clinics and the respondents' own homes, and transport.

Agents
Two types of agents were simulated in the model, people and households.

Households
Households were simulated as agents, for the purpose of grouping people into households with the desired size distributions. Households had the following state variables: Other temporary household-level state variables were used to store information on the disease states of household members when estimating transmission probabilities in the household (see section 'Mtb transmission -Household members')

Household sizes
Empirical data were available from the study population on the number of people aged 15+ years in each household. An exponential distribution was fitted to data on the cumulative proportion of households below each size, and the distribution was sampled from and rounded up to the nearest whole number to create desired household sizes in the model ( Figure S1). Mean household sizes were similar between the model and the empirical data both from the perspective of households (model=3.64, data=3.97), and from the perspective of individuals (model=6.75, data=6.55).

Model initialisation
To initialise the model, N empty households were created, where N = round (10,000 /mean_hh_size). Each empty household sampled a household size from an exponential distribution, rounding up to the nearest whole number (see section 'Household sizes' for details), and then created that number of people to populate the household. This gave an initial population size of approximately 10,000.
The newly created people were each assigned a sex, with a probability of 0.5 of being male and 0.5 of being female, and a clinic_group with and a probability of 0.5 of being 'high' and 0.5 of being 'low'. They were then assigned an age_group, with probabilities assigned by input parameters, and varying by sex; and an age, drawn from a uniform distribution between the minimum and maximum ages in their age_group. A random infection_seed_proportion were seeded with latent infection, with no risk of progression without reinfection. A random tb_seed_proportion were then seeded with TB disease, with probability prop_smearpos_HIV0 becoming smear+ and the rest smear-.
The model was run with a constant population size for 100 years, and then a further 100 years with a growing population size, to allow the population age distribution and TB incidence and mortality to reach equilibrium. At that point, the model was considered to represent the year 2000, and realistic trends in HIV and TB were simulated from that point onwards.

Model scheduling
The majority of events in the model were simulated using continuous time.
The two exceptions to this were the creation of new people, and the Mtb transmission process, which used a monthly time step.

Demography
Individuals were introduced into the model at age 15. People aged <15 were not modelled, as the risk of Mtb transmission from children is low 2 , and contact data were not available from children from the study population.
During the initial run-in period, a constant population size of 10,000 was simulated. Each month, the number of people alive in the model was counted, and additional people created to restore the model population size to 10,000. After the initial run-in period, a constant birth rate per person alive was simulated, with the number of new people to be created each month equal to binomial(population size, birth_rate). The background morality rates varied by age and sex, and were constant over time within each age group. TB and HIV mortality are described in the sections on TB and HIV.

Fitting targets
The model was fitted to provincial-level data from KwaZulu-Natal on the estimated growth in population size between 2015 and 2019, the proportion of the population who are male in 2018, and the proportion of men and women in each of the three simulated age groups, by varying the simulated birth rate and age and sex specific background mortality rates. As in-and out-migration were not explicitly simulated, the background mortality rates were not designed to accurately reflect true (non-HIV and non-TB) mortality rates by age, but instead to also incorporate the effects of inand out-migration on the population age distribution.

Social contact
Three types of social contact were simulated in the model: contact between household members, contact occurring in clinics, and contact occurring in all other locations.

Household members
In the model, it was assumed that each individual has contact_time_each_hh_mem = 572 hours of indoor contact with each member of their household each month (18.8 hours per day * 365.25 days / 12 months).

Clinics
In line with the empirical data, the rate of clinic visiting in the model varied by sex and HIV/ART status, but not by age group. For each sex and HIV/ART status strata, 50% of the simulated population was assumed to be in a high clinic visiting group, and 50% in a low clinic visiting group.
Clinic visiting rates in each group, for each strata, were determined by fitting a Poisson distribution to the data on the proportion of people in each strata who visited a clinic 0, 1 ,2-5 or 6+ times in the past six months, and the overall rate of clinic visiting in the strata, using a sum of least squares approach. Individuals changed between the high and low clinic visiting groups every six months with probability clinic_rate_switch_prob.
The rate of clinic visiting also varied for individuals with untreated TB disease (in the states smearpositive disease (smear+) and smear-negative disease (smear-)). Compared to individuals of the same sex, HIV/ART strata, and clinic visiting group, the rate of clinic visiting in people with untreated TB disease was increased by a factor of increased_contact_time_clinics_tb.
It was assumed in the model that all individuals had 140 adult contact hours on each clinic visit.
Individual clinic visits were not explicitly simulated in the model, instead each individual had a set amount of contact time in clinics each month (e.g. contact_time_clinic_m_HIV01_low), equal to the assumed mean number of clinic visits in a month (by sex, HIV/ART status, and clinic visiting group) multiplied by the mean contact time per visit.   waiting areas show large amount of variation in ventilation rates between different spaces, but they suggest that clinic spaces are generally better ventilated on average than people's homes 5 . We assumed in the model that the rate of transmission from a person with TB disease to a person without is 2.8 times higher in homes than in clinics. As the model is calibrated to an estimate of the proportion of disease that results from transmission between household members, however, the assumption made about ventilation rates in homes vs other spaces has little effect on the results (see Section 2.9.8).

Other locations
Limited data were available on ventilation rates from other types of location, and showed large amounts of variation 6 . Nevertheless, rates for most locations were more in line with the higher ventilation rates found in clinic waiting areas than the lower rates found in people's houses. For this reason, we assumed in the main scenario in the model that the rate of transmission between a person with TB disease and a person without is the same in other locations as in clinics.
The effects of the assumptions made about ventilation rates in clinics and other locations were explored in a sensitivity analysis (See section 2.14 Uncertainty analysis).  Resistance type in the model effected the TB treatment duration. The treatment duration for non-MDR-TB was always six months. For MDR-TB, it was 24 months for all people starting TB treatment before 2016, then 24 months with probability 0.3, and 11 months with probability 0.7 7 8 .

Disease progression
The rate of developing tuberculosis disease following infection depended on an individual's time HIV-, HIV+ART-, and HIV+ART+ people with TB disease self-cured at rate self_cure_rate_HIV0, self_ cure_rate_HIV1, and self_cure_rate_HIV2 respectively. Upon self-cure, individuals re-entered the latent stage, resetting their time since infection back to zero.

Treatment
Individuals with TB started treatment each month with probability treatment_rate_HIV0 if HIV-, and treatment_rate_HIV12 if HIV+. These rates took the value treatment_rate_HIV0_early and treatment_rate_HIV12_early respectively before treatment_rate_change_year, and treatment_rate_HIV0_late and treatment_rate_HIV12_late respectively afterwards.
After the year that ART was first introduced into the model, ART_intro_year, upon starting TB treatment, all HIV+ART-people became HIV+ART+.
Treatment lasted for treatment_duration_DS months if non-MDR-TB, and treatment_duration_MDR months if MDR-TB. Individuals successfully finishing treatment re-entered the latent stage. Upon doing so, they reset their time since infection back to zero, reflecting the high rates of disease recurrence following treatment 9 10 .
Individuals receiving TB treatment dropped out of treatment each month with probability TB_treatment_dropout_rate_ DS if they had non-MDR-TB and, TB_treatment_dropout_rate_MDR if they had MDR-TB. Upon dropping out of treatment, they returned to active TB disease, with the BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s)

Mortality
TB mortality rates in the model depended on disease type (smear-or smear+), HIV/ART status, and whether someone was receiving treatment or not.
When on treatment, the annual TB mortality rate was TB_mortality_rate_treatment_DS for people with non-MDR-TB, and TB_mortality_rate_treatment for people with MDR-TB. Different TB mortality rates by HIV status while on TB treatment were not simulated, as empirical data showed little difference in treatment success by HIV status in South Africa 11 .

Prevalence of infection in 15-year olds
In 2013, 14.4% of 6-8 year olds were found to be infected with Mtb or to be on TB treatment in KwaZulu-Natal, giving an estimated annual rate of infection rate 2.1% 12 . Adjusting by reductions in estimated TB incidence between 2013 and 2018, and by increases in attack rates between childhood and adolescence 13 , we estimated that around 24.2% of adolescents in KwaZulu-Natal in 2018 were infected with Mtb. Upon being created at the age of 15 years, people in the model therefore set their state to latent with probability 0.242. The remaining people were assumed to be uninfected.
In calculating rates of progression to active disease in individuals with Mtb infections at the point of their creation at age 15 in the model, we assigned them a time of infection, time_of_infection, from a uniform distribution covering the 15 years before their creation. Their rate of disease progression was then calculated using the same method as was used for people infected at ages >15 years.
Progression to disease that occurred prior to the age of 15 was not included in the model.

Changes in TB natural history parameters over time
To reflect secular trends not captured by other time varying parameters in the model (for instance, improvements in nutrition and housing), a step change was modelled in TB_parameter_change_year. In TB_parameter_change_year, the simulated rate of Mtb transmission (transmission_prob), and the simulated rates of progression to TB disease following infection were reduced by a factor of decreased_tb_rates_late.

Fitting targets
The model was fitted to a range of TB incidence, mortality, and treatment outcome estimates (see section 'Modelling fitting targets').
The model was also fitted to the central value of a range of estimates for the proportion of disease that results from transmission between household members in sub-Saharan African countries 14 . This was done by varying the degree of individual variation in infectiousness between people with tuberculosis, with higher levels of variation leading to a lower proportion of disease resulting from transmission between household members.

Mtb transmission
Mtb transmission in the model was scheduled on a monthly time step. Three transmission 'locations' were simulated, with transmission in each location simulated in turn each month: transmission between household members, transmission in clinics, and transmission in other indoor locations (including transport). Random mixing was assumed in clinics and in other locations.
In all locations, the parameter transmission_prob determined the baseline probability of transmission per minute contact between each uninfected or latent person and each person with smear+ or smear-TB. transmission_prob took the value transmission_prob_early before TB_parameter_change_year, and transmission_prob_early * decreased_tb_rates_late afterwards.
The baseline transmission_prob was then adjusted for a number of factors: • The simulated ventilation level in the location. The effect of ventilation levels on the rate of transmission is described in the section 'Ventilation'.
• The smear status of the person with TB. We assumed that people with smear-disease are 78% less infectious than people with smear+ TB 15 . • Whether the exposed person was uninfected or latent, and their HIV/ART status. We assumed that latent infection provides 72% protection against reinfection in HIV-people 16   The infectiousness parameter was assumed to incorporate the effects of all factors that have an effect on the infectiousness of a person with TB, with the exception of whether the disease is smear+ or smear-.

Individual-level variation in infectiousness
Individual-level variation in infectiousness was simulated when determining Mtb transmission between household members, because the variation acts to reduce the rate of transmission between highly regular contacts such as household members, through increasing the effects of saturation 14 . Not incorporating this variation would therefore have resulted in an unrealistically high proportion of disease in the model coming from transmission between household members.
Individual-level variation in infectiousness was not used in the model when determining Mtb transmission in clinics and other locations. Instead, the overall mean value of infectiousness, 1, was used for all people. This reduced model stochasticity, speeding up the model fitting process, and meaning that far fewer model runs needed to be done per final scenario and intervention. As random mixing was simulated in both clinics and other locations, this had no effect on the average proportion of disease that results from transmission in clinics and other locations in the model. For each susceptible or latent individual in the household, the probability of infection each month was calculated as:

Household members
Where: • reinfection_relative_risk = 1 if the individual was uninfected, reinfection_relative_risk_HIV0 if they were HIV-and latently infected, reinfection_relative_risk_HIV1 if they were HIV+ARTand latently infected, and reinfection_relative_risk_HIV2 if they were HIV+ART+ and latently infected • Ws = 1 when s = 1, and Ws = reduced_transmission_smearneg when s = 0 The probability that people infected with Mtb from transmission from a household member were infected with an MDR strain was calculated as:

Clinics
Each month, the total contact number of people in each class was counted, with class defined as the 60 strata generated by all combinations of: • Sex (male, female) • HIV/ART status (HIV-, HIV+ART-, HIV+ART+) • Clinic visiting group (high, low) • For each susceptible or latent individual in the model, the probability of infection each month from transmission in clinics was then calculated as: Where: • reinfection_relative_risk = 1 if the individual was uninfected, reinfection_relative_risk_HIV0 if they wer HIV-and latently infected, reinfection_relative_risk_HIV1 if they were HIV+ARTand latently infected, and reinfection_relative_risk_HIV2 if they were HIV+ART+ and latently infected.
• contact_time_clinics was equal to the mean monthly contact time in clinics for someone of the individual's class. The probability that people infected with Mtb from transmission in clinics were infected with an MDR strain was calculated as:

Other locations
Each month, the total contact number of people in each class was counted, with class is defined as the 90 strata generated by all combinations of: For each susceptible or latent individual in the model, the probability of infection each month from transmission in other locations was then calculated as: Where: • reinfection_relative_risk = 1 if the individual was uninfected, reinfection_relative_risk_HIV0 if they were HIV-and latently infected, reinfection_relative_risk_HIV1 if they were HIV+ARTand latently infected, and reinfection_relative_risk_HIV2 if they were HIV+ART+ and latently infected.
• Ws = 1 when s = 1, and Ws = reduced_transmission_smearneg when s = 0 • contact_time_other was equal to the mean monthly contact time in other locations for someone of the individual's class.
The probability that people infected with Mtb from transmission in other locations were infected with an MDR strain was calculated as:

HIV/ART
Three HIV states were simulated in the model: HIV-, HIV+ART-, and HIV+ART+. To capture changes in estimated and projected HIV prevalence over time, the value of the HIV incidence parameters for each age group and sex changed twice in the model, in From ART_start_rate_change_year, all HIV+ART-people starting TB treatment were made HIV+ART+.
HIV mortality was simulated as a constant rate of (non-TB) HIV-related mortality for all HIV+ARTpeople (HIV1_mortality_rate), and all HIV+ART+ people (HIV2_mortality_rate),

Effects on TB
HIV and ART status effected a number of TB-related rates and probabilities in the model: • TB mortality rates • Rates of progression to disease

Changes in HIV parameters over time
CD4 counts for HIV+ART-people were not explicitly simulated, with HIV+ART-being simulated as a single, homogenous group, varying only with age group and sex. As ART coverage increased over time in South Africa, however, the average CD4 count of people not on ART is likely to have risen, and the impact on TB natural history of being HIV+ART-is likely to changed. To allow the effects of increased ART coverage on TB to be adequately captured in the model, enabling the model to be fitted to trends in TB incidence over time, a step change in the values of certain HIV related parameters was simulated, starting in change_HIV1_parameters_year.
From change_HIV1_parameters_year, the degree of protection that latent infection gave against reinfection in HIV+ART-people, reinfection_relative_risk_HIV1, was increased from reinfection_relative_risk_HIV1_early to reinfection_relative_risk_HIV1_late, and the rate of developing disease in more than one year following infection in HIV+ART-people was decreased from develop_tb_reactivation_rate_HIV1_early to develop_tb_reactivation_rate_HIV1_late. From TB_parameter_change_year, these rates were also decreased by decreased_tb_rates_late (see section 'Changes in TB natural history parameters over time' and Table S6). As the rate of developing disease in the first year following infection in HIV+ART-people was calculated relative to the rate in subsequent years in the model, this also decreased the rate in the first year following infection.

Fitting targets
The model was fitted to a range of HIV prevalence and ART coverage targets, based on empirical estimates from the study population 17 . These are described in full in the section 'Fitting targets'.
In addition to this, the model was fitted to estimated future trends in HIV prevalence and ART coverage by sex, from provincial HIV model (Thembisa) estimates 18 19 . As the Thembisa estimates were for the province as a whole, and the model was fitted to historic trends from the study population, the model was fitted to estimates changes in HIV prevalence and ART coverage by sex between 2020 and 2030, rather than the absolute estimates. 3) Ultraviolet Germicidal Irradiation (UVGI) system. We assumed in this intervention that appropriate and well maintained UVGI systems are installed in all indoor clinic waiting areas. This was implemented in the within-clinics model through an additional quanta clearance rate, equivalent to a ventilation rate of 24 ACH (95% CI 9.9-62) 24 .

4) Surgical mask wearing by patients.
We simulated a scenario where 70% of patients wear surgical masks 90% of the time. Masks were assumed in the within-clinics model to reduce the rate of quanta production by 75% (95% CI 56-85%) 25 , and have no effect on rate of infection for the person wearing the mask 26 .  27 . This means that they do not need to queue at clinics unnecessarily. The purpose of this intervention was to increase the utilisation of CCMDD and similar programmes by eligible patients, and to ensure that pick-up points do not require patients to queue at clinics. We assumed that 92% (95% CI 84-95%) of patients could have their ART appointments reduced to once every 6 months 28 , and that the remaining 8% of people need monthly ART appointments. This was implemented in the withinclinics model through removing 31% (IQR 22-34%) of ART patients, chosen at random each model run.
6) Queue management system with outdoor waiting areas. Empirical data show that clinic waiting areas are often crowded, and that in many clinics patients wait in unsuitable areas such as corridors 29 . This is partly due to patient concerns that if they wait in other areas, they may not hear their name being called, and may miss their turn. This intervention therefore combined a large, covered outdoor waiting area with a queue management system, such as numbered tickets or an electronic tracking system. We  Table S7).
The 'best estimates' of intervention effects in this model were informed by the median impacts from the within-clinics model. The minimum and maximum estimates were informed by the interquartile ranges from the within-clinics model. The interquartile range was used, rather than the full range, as the most extreme effects from the within-clinics model were assumed to reflect day to day variation, rather than genuine uncertainty in intervention effects.

Results calculations
When calculating the proportion of disease that resulted from transmission in clinics in the model, simulated individuals who developed disease from an infection that occurred before the age of 15 years were not included, as their location of infection could not be determined.
Intervention effects on TB incidence and mortality were calculated as relative changes in rates, compared to a scenario where no interventions are simulated. As the simulated proportion of people created in the model at age 15 years who had a latent infection is constant over time, simulated individuals who developed or died from TB disease from an infection that occurred before the age of 15 years were not included when estimating intervention effects on TB incidence and mortality.

Uncertainty analyses
A number of univariate sensitivity analyses were conducted: • Proportion of outside-household contact time occurring in clinics (clinic contact time).
From the social contact data, overall, we estimated that 5.3% (95% CI 2.  Table S8. • Prevalence of TB in clinic attendees relative to the general population (TB in clinics). In the main scenario, the model was fitted to a prevalence of TB in clinic attendees relative to the community prevalence of 1.86. In the sensitivity analysis, the model is fit to the upper bounds of the empirical 95% confidence interval (1.1-3.1) 3 . Fitting to the lower bound would have required the value of increased_contact_time_clinics_tb to be less than one. In other words, it would have required simulating a lower rate of clinic visiting in people with TB compared to people without, controlling for sex and HIV/ART status. This was considered to be implausible, therefore increased_contact_time_clinics_tb = 1 was used as the lower bound.
• Proportion of disease from household transmission (Household transmission). In the main scenario, we fitted the model to 13.5% of disease resulting from transmission between household members. In the sensitivity analysis, the model was fitted to 8% and 19% 14 of disease resulting from transmission between household members. This was achieved primarily by changing the value of infectiousness_var. • Movement between high and low clinic visiting groups (Clinic risk groups). As the social contact survey collected data on number of clinics visits over a six-month period only, we were unable to distinguish the extent that differences in clinic visiting rates between people of the same sex and HIV/ART status were due to long-term, stable differences vs shorter term fluctuations in clinic use. In the main scenario, we simulated people switching between clinic visiting risk groups every six months with probability clinic_rate_switch_prob = 0.25. In the sensitivity analysis, we simulated people switching with probability 0 and 0.5.

• Clinic visiting rates by HIV+ART-people, relative to HIV-people (HIV+ART-clinic visiting).
In the social contact data collection, only 13 people reported being HIV+ART-. In addition, HIVstatus was self-reported, and we could therefore not accurately distinguish between HIVand undiagnosed HIV+ people, particularly when the reported date of the last HIV-test was not recent. We therefore had no empirical data on rates of clinic visiting in HIV+ART-people.
In the main scenario, we assumed that the rates are the same in HIV+ART-and in HIVpeople, and determined the rates from the empirical data for all people who did not report being on ART. In the sensitivity analysis, we assumed that rates in HIV+ART-people are half that of HIV-people, and that rates in HIV+ART-people are the same as for HIV+ART+ people.
In both scenarios, we also adjusted the HIV-clinic visiting rates to keep the overall mean clinic visiting rates in 2020 for HIV+ART-and HIV-people constant. The simulated clinic contact times are shown in Table S8.
• Future HIV incidence. Estimated future trends in HIV incidence were taken from the projections from a provincial-level HIV model, Thembisa 18 19 , with the model fitted to the estimated change in HIV prevalence in men and women between 2020 and 2030. While Thembisa did provide 95% limits for its estimates, we considered them to be unrealistically narrow. For instance, the 95% limits for the projected prevalence of HIV in men aged  in 2030 was 11.4-12.3%. In the sensitivity analysis, we therefore chose to simulate relative changes in HIV incidence by sex from 2020 compared to the preceding time period, that were 50% lower and 150% higher than the simulated changes in the main scenario.
In all sensitivity analyses, the model was recalibrated to the same fitting targets (with the exception of the targets explicitly changed in the sensitivity analysis).         TB_mortality_rate_treatment_ MDR 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14 0.14  Table S12. Model fit to fitting targets, in the best estimate scenario and sensitivity analysis scenarios. 'Low' and 'high' refer to changes that decrease and increase the proportion of disease that results from transmission in clinics respectively. *Indicates fitting outputs where the target value was changed in the sensitivity analysis.      No data were available on excess tuberculosis risk in clinic staff in our study setting. A recent systematic review of TB incidence in healthcare workers estimated that the ratio of the rate of TB in healthcare workers compared to the general population in high TB burden settings was 4.32 (95% CI 2.36-7.91 22 ).