We selected the 100 most populous municipalities in Brazil (on the basis of mean population between 2008–2017) plus Palmas, the 1 state capital that was not among those 100. We obtained TB treatment notifications from Brazil’s National Notifiable Disease Information System (SINAN) (5) and death data from the Mortality Information System (SIM) (6), representing 438,163 notified TB cases and 45,984 TB-related deaths. Using these data, we estimated a Bayesian model of tuberculosis incidence (M.H. Chitwood et al., unpub. data, https://doi.org/10.2139/ssrn.3463278) in which incidence is approximated by the sum of 3 numbers: treatment initiations, deaths before treatment initiation, and disease resolutions before treatment initiation for a municipality in a given year. We reported the annual incidence rate as absolute incidence divided by population size and the fraction receiving treatment (fraction treated) as the number initiating treatment divided by incidence in a given year. The fraction treated differs from the case detection rate by considering loss to follow-up between diagnosis and treatment as an additional mechanism contributing to undertreatment. We also estimated the incidence of untreated TB (untreated TB rate) as the product (incidence rate) × (1 − fraction treated), to produce a combined measure of elevated incidence and inadequate case detection.

Figure 1

Figure 1. Modeled tuberculosis (TB) burden in 101 largest municipalities and state capitals of Brazil, 2017. Gray curves indicate isopleths of untreated TB incidence × (1 − fraction treated). Municipalities in the...

Across all 101 municipalities in 2017, there were 53.2 treatment notifications/100,000 population; we estimate a TB incidence rate of 58.6 (range 11.6–169) cases/100,000 population (Table). In 2017 São Vicente had the highest estimated TB incidence, 169 (95% CI 154–185) cases/100,000 population, and Palmas had the lowest, 11.6 (95% CI 9.3–14.3) cases/100,000 population. We estimate that the fraction treated ranged from 0.778 (95% CI 0.687–0.852) to 0.969 (95% CI 0.934–0.990)/100,000 population and the untreated TB rate ranged from 0.723 (95% CI 0.231–1.61) to 23.0 (95% CI 15.1–34.8)/100,000 population (Figure 1).

During 2008–2017, there were 438,163 TB treatment notifications; for this period we estimate that there were 488,329 (95% CI 474,715–507,676) incident TB cases, of which 49,778 (95% CI 36,072–69,217) did not initiate treatment. We observed a decrease in notifications from 56.6/100,000 population in 2008 to 53.2/100,000 population in 2017; over this period we estimate that average incidence decreased from 63.9 (range 13.7–138) to 58.6 (range 11.6–169)/100,000 population. Incidence decreased at an average annual rate of 0.95% (range −5.41% to 4.73%), the fraction treated increased at an average annual rate of 0.290% (range −0.966% to 3.55%), and the untreated TB rate decreased at an average annual rate of 2.88% (range −17.4% to 7.98%).

Figure 2

Figure 2. Municipalities of Brazil with the greatest decreases and increases in untreated tuberculosis (TB), 2008–2017. A, B) The 10 municipalities with the greatest decrease in untreated TB, showing the difference between...

We compared the 10 municipalities with the largest absolute decrease and the 10 with the largest absolute increase in the untreated TB rate (Figure 2). In the municipalities with the largest decrease in untreated TB, the fraction of treated TB cases increased at an average annual rate of 1.23% (0.619–2.17), and incidence decreased at an average annual rate of 1.31% (−3.16 to 2.31) (Figure 2, panels A, B). We estimated that incidence increased in 2/10 municipalities, most notably São Vicente, which had an average annual rate of increase of 2.31% (95% CI 0.642%–3.89%).

In the 10 municipalities with the largest increase in untreated TB, the fraction treated decreased; average annual rate was 0.596% (0.252–0.985) and average incidence rate increased (0.732%; range −2.82 to 3.62) (Figure 2, panels C, D). Although the fraction treated decreased on average, CIs were wide and crossed 0 for the majority of estimates. The change in incidence was heterogenous in this group, ranging from an average decrease of 2.83% (95% CI 1.75%–3.93%) per year in Duque de Caxias to an average increase of 3.63% (95% CI 1.82%–5.35%) per year in Campos dos Goytacazes.

Using a recently developed Bayesian approach for subnational TB estimation (M.H. Chitwood et al., unpub. data), we estimated the TB incidence rate, fraction treated, and the untreated TB rate for 101 large municipalities in Brazil during 2008–2017. We found that the incidence rate decreased on average and the fraction treated increased on average over the study period. However, in several high-burden municipalities, TB incidence rose and the fraction treated declined, increasing the untreated-TB rate and indicating gaps in local TB control efforts. Comparing our results with a similar state-level analysis of TB trends in Brazil, we found that large municipalities are more heterogenous and have more volatile trends in incidence and fraction treated than states.

The rate of untreated TB communicates both the size of the epidemic and the strength of the response. Municipalities with the highest incidence or the lowest fraction treated may not be the same municipalities with the highest untreated TB rate; an area with a moderate TB incidence and a moderate fraction treated could have a nontrivial rate of untreated TB. If municipalities in need of additional programmatic support were identified based only on the estimated incidence or fraction treated, cities with moderate incidence may be overlooked.

Because we applied a common set of assumptions across all municipalities, our approach may not account for local factors that influence the ratio between reported TB cases and deaths attributed to TB. In our analysis, this ratio provides a signal of the completeness of case detection. If TB death reporting in a municipality were biased downwards (e.g., many TB deaths were misattributed to other causes), the result would be an upward bias in the estimate of the fraction of cases treated. We assume that differences between deaths of persons who have initiated treatment and deaths reported in SIM are due to deaths that occur before treatment. A records linkage of SIM and SINAN was not possible for this analysis. Such a linkage would enable more precise quantification of the frequency of death before treatment initiation. If the overlap between the systems was lower than expected (e.g., more deaths before treatment initiation), our model would underestimate TB burden.

In this analysis, we identified municipalities, such as São Vicente, in which both the fraction treated and incidence increased on average. If these estimates are correct, our findings suggest that factors other than treatment coverage, such as delays between disease onset and treatment initiation, low treatment completion rates, or worsening nutrition and housing quality, could be driving trends in TB incidence. Further analysis of municipalities with both increasing fraction treated and increasing incidence is warranted to elucidate which factors drive increasing TB incidence despite improvements in treatment coverage.