Systematic differences in TB treatment outcomes across Brazil by patient- and area-related factors: an analysis of national disease registry data

Abstract

Background

Many individuals initiating tuberculosis (TB) treatment do not successfully complete the regimen. Understanding variation in treatment outcomes could reveal opportunities to improve the effectiveness of TB treatment services.

Methods

We extracted data on treatment outcomes and patient covariates from Brazil’s National Disease Notification Information System, for new TB patients diagnosed during 2015–2018. We analysed whether or not patients experienced an unsuccessful treatment outcome (any death on treatment, loss to follow-up or treatment failure). We constructed a statistical model (logistic regression with regularised two-way interactions) to predict treatment outcomes as a function of socio-demographic factors, co-prevalent health conditions, health behaviours, membership of vulnerable populations and form of TB disease. We used this model to decompose state- and municipality-level variation in treatment outcomes into differences attributable to patient-level and area-level factors.

Results

Treatment outcomes data for 259 449 individuals were used for the analysis. Across Brazilian states, variation in unsuccessful treatment due to patient-level factors was substantially less than variation due to area-level factors, with the difference between best and worst performing states (lowest and highest fraction with unsuccessful treatment, respectively) equal to 7.1 and 13.3 percentage points for patient-level and area-level factors. Similar results were estimated at the municipality level, with 9.3 percentage points separating best and worst performing municipalities according to patient-level factors, and 20.5 percentage points separating best and worst performing municipalities to area-level factors. Results were similar when we analysed loss to follow-up as an outcome.

Conclusions

Our analysis revealed substantial variation in TB treatment outcomes across states and municipalities, with only a minority attributable to patient-level factors. Area-level variation likely reflects consequences of differences in health system organisation or socio-environmental factors not reflected in patient-level data. Further research on these factors is needed to identify effective approaches to TB care, reduce geographic disparities and improve treatment outcome.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Understanding the factors that determine tuberculosis (TB) treatment outcomes is important for improving the effectiveness of TB services. Studies of TB treatment outcomes in Brazil have revealed substantial differences in outcomes between states, even after controlling for patient-level factors. Subnational differences in treatment outcomes could reflect modifiable features of health service organisation or practices, which could be targeted to improve overall programme outcomes and reduce location-based disparities.

WHAT THIS STUDY ADDS

• This study quantified differences in TB treatment outcomes at state- and municipal-level in Brazil, adjusting for patient-level factors to allow fair comparisons between geographic areas. The results showed that area-level variation in treatment outcomes (ie, variation that could not be explained by patient-level factors) was substantial. For example, the difference between best and worst-performing states attributable to area-level factors was 13.3 percentage points, in comparison to a 7.1 percentage point difference between best and worst-performing states attributable to patient-level factors, and similar results were found in a municipal-level analysis.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Substantial area-level differences in TB treatment outcomes across Brazil highlight the potential impact of health system organisation, clinical practices and local socioeconomic factors in determining the success of TB treatment. This analysis identifies states and municipalities that represent positive and negative outliers in terms of treatment outcomes, which could be studied to identify factors associated with high- and low-performing services. In addition, high-performing areas provide examples of successful practices that could be adopted in other areas, while low-performing areas could be provided additional support to improve outcomes.

Background

Tuberculosis (TB) is one of the leading causes of infectious disease deaths, killing over 1 million individuals each year.¹ For individuals with TB who are unable to access treatment, approximately half will die.² Even for individuals who start TB treatment, a substantial fraction die or discontinue treatment before completing the regimen, or are recorded as having evidence of treatment failure. Of the 7.3 million individuals diagnosed in 2022 for whom TB treatment data were reported to WHO, 12% experienced an unsuccessful treatment outcome.³ In addition to the poor health outcomes faced by these patients, loss to follow-up and treatment failure create additional opportunities for onward transmission, as well as the development and propagation of TB drug resistance.

Understanding the factors that determine TB treatment outcomes and addressing barriers to high-quality care are important steps for improving the effectiveness of TB services. Analysis of treatment outcomes and other surveillance data has been promoted to identify gaps in TB care cascades, reveal variation in outcomes and target programme action towards areas with the greatest opportunities for improvement.^{4 5} In a number of high TB burden countries, routinely collected TB outcomes data have been used to quantify subnational differences in the fraction of patients successfully completing treatment.⁶,¹²

In Brazil, TB epidemiology and health services have been shown to vary across the country, with substantial differences in TB incidence, mortality and case detection rates reported at state and municipal level.¹³,¹⁸ Similarly, subnational differences have also been reported for TB treatment outcomes.¹⁹,²¹ Some part of this variation will be explained by differences in individual-level TB risk determinants, and several studies have reported systematic differences in risks of treatment mortality and loss to follow-up according to individual demographics, health-related behaviours and co-prevalent medical conditions.²¹,²⁴ However, those studies that have simultaneously considered both patient-level and spatial factors have found substantial residual variation even after patient-level factors are considered,^{20 21} with one study finding the odds of unsuccessful treatment (controlling for patient-level factors) to vary by a factor of 2.9 between best and worst performing states.²¹

While some patient-level differences in treatment outcomes will reflect causal processes that are difficult to address (eg, the elevated TB case fatality associated with advanced age), differences in average treatment outcomes between geographic areas could reflect modifiable features of health service organisation or practices, which could be targeted to improve overall programme outcomes and reduce location-based disparities. Similarly, areas with above-average outcomes could provide examples of successful practices that could be promoted for broader adoption. In this study, we conducted an analysis to systematically quantify subnational differences in TB treatment outcomes in Brazil, adjusted for the range of patient-level factors found to be associated with treatment outcomes in past studies. Based on this analysis, we decompose overall state- and municipality-level differences in treatment outcomes to describe the components attributable to patient-level and area-level factors and identify the highest- and lowest-performing locations according to each of these metrics.

Methods

Study population and data

The study population represented persons without a prior history of TB who were diagnosed with TB and initiated on treatment in Brazil during the years 2015–2018. We obtained data on this study population from Brazil’s National Disease Notification Information System (SINAN).²⁵ For each individual, we extracted variables describing features of disease presentation and treatment, demographic and socioeconomic factors, the presence of co-morbidities and health-related behaviours. We assigned patients to municipalities based on their recorded municipality of residence at the date of TB notification. For individuals for whom the residential location was not recorded in SINAN (either missing or could not be linked to a municipality), we used the municipality where the individual received treatment as a proxy for their residence.

Definition of unsuccessful treatment outcome

We analysed a binary outcome indicating whether individuals in the study population experienced an unsuccessful treatment outcome. Individuals were categorised as having an unsuccessful treatment outcome if they were recorded as having died from TB or other cause during TB treatment, experienced regimen failure (positive sputum smear or culture in the fourth month or two consecutive months after the fourth month of treatment initiation) or were lost to follow-up (non-attendance at scheduled treatment visits for 30 or more days after treatment initiation). Individuals recorded as experiencing treatment success represented individuals completing the treatment regimen without evidence of regimen failure. We excluded individuals who could not be assigned to unsuccessful or successful treatment outcome based on these definitions. In secondary analyses, we examined a binary indicator for whether or not the individual was lost to follow-up, using the same study sample as the main analysis.

Adjustment for patient-level factors

We constructed statistical models predicting the treatment outcome as a function of patient-level factors. These included variables previously associated with differences in TB treatment outcomes in this setting: sociodemographic factors (sex, age group, educational level, self-declared race), co-prevalent health conditions (HIV, diabetes), health behaviours (illicit drug use, alcohol use, smoking), membership of vulnerable populations (incarcerated, homeless, immigrant) and diagnosed form of TB disease (pulmonary, extrapulmonary or both).²¹

We trained and evaluated a range of prediction models based on linear regression, logistic regression, boosted regression trees and ridge regression. Each of these models was constructed for the binary treatment outcome. We assessed out-of-sample predictive performance via 10-fold cross-validation and compared models in terms of the Brier score, C-statistic, visual inspection of calibration plots, calibration intercept and slope and Hosmer-Lemeshow goodness-of-fit statistic. We selected the best-fitting model as the one that minimised overall predictive error (assessed via the Brier score). As a final validation step, we re-fit this model to data for the years 2015–2017 and assessed how well this model predicted 2018 treatment outcomes data not used for model fitting. The final predictive model was constructed by fitting the chosen model to all data (2015–2018) and was used for all subsequent analyses. With this final model, we estimated the probability of an unsuccessful treatment outcome for each individual in the study cohort.

Estimation of state-level outcomes

We decomposed inter-state variation in the fraction of patients experiencing an unsuccessful treatment outcome into a component attributable to patient-level factors and a second component attributable to area-level factors. We defined $E_i^{All}$ as the excess risk of unsuccessful treatment for state i compared with the national mean, $E_i^{Patient}$ as the excess risk of unsuccessful treatment attributable to patient-level factors and $E_i^{Area}$ as the excess risk of unsuccessful treatment attributable to area-level factors, with $E_i^{All}$ equal to the sum of $E_i^{Patient}$ and $E_i^{Area}$. To estimate each of these outcomes, we calculated:

$$E_i^{All}=\frac{u_i}{n_i}-\frac{{\sum }_i{u}_i}{{\sum }_i{n}_i}$$ (1)

$${\displaystyle {\displaystyle {\displaystyle E_i^{Patient}=\frac{{\hat{u}}_i}{n_i}-\frac{\sum _i{u}_i}{\sum _i{n}_i}}}}$$ (2)

$${\displaystyle {\displaystyle E_i^{Area}=\frac{u_i}{n_i}-\frac{{\hat{u}}_i}{n_i}}}$$ (3)

where $u_i$ and $n_i$ represent the observed number of unsuccessful treatment outcomes and total number of observations for state i over the study period, respectively, and ${\widehat{u}}_i$ represents the predicted number of unsuccessful treatment outcomes for the same period, calculated by summing the probabilities generated by the predictive model. Based on these calculations, $E_i^{Patient}$ represents the difference in treatment outcomes in a given state, as compared with the national average, that is produced by differences in patient case-mix, while $E_i^{Area}$ represents the residual variation in treatment outcomes not explained by differences in case-mix. We multiplied the values for $E_i^{All}$, $E_i^{Patient}$ and $E_i^{Area}$ by 100, so they represented percentage point differences in the proportion of individuals experiencing unsuccessful treatment outcomes and used these results to describe the relative contribution of patient-level and area-level factors to overall differences in treatment outcomes across states.

Estimation of municipality-level outcomes

We modified this approach to estimate $E_i^{All}$, $E_i^{Patient}$ and $E_i^{Area}$ for municipalities. As some municipalities have small numbers of treated patients, using the same approach as used for states could result in noisy estimates of $E_i^{Area}$. Instead, we fit a binomial random effects model for the fraction of patients experiencing unsuccessful treatment in each municipality, with the predicted fraction experiencing unsuccessful treatment ($\frac{{\hat{u}}_i}{n_i}$) used as an offset term. Based on this specification, the fitted random effect term for each municipality represented the estimated difference in municipality-level outcomes due to area-level factors ($E_i^{Area}$). For municipalities with small numbers of cases, the random effect specification pooled this estimate towards the overall mean, while for municipalities with large numbers of cases, this pooling effect was negligible. $E_i^{Patient}$ was calculated in the same way as described for states, and $E_i^{All}$ was calculated as the sum of $E_i^{Patient}$ and $E_i^{Area}$. We used the municipality-level results to describe spatial patterns in $E_i^{Patient}$ and $E_i^{Area}$ across Brazil.

Uncertainty analysis and secondary outcomes

We quantified uncertainty in the estimates for each outcome by recalculating results for 10 000 bootstrap samples of the underlying patient data and estimated 95% uncertainty intervals from the 2.5th and 97.5th percentiles of the distribution of bootstrap results for each outcome.

We repeated these analytic steps for an alternative outcome defined as the fraction of TB patients lost to follow-up, with the rationale that this may be the contributor to unsuccessful treatment outcomes that is more amenable to programme action.

Patient and public involvement

This study was undertaken as a retrospective analysis of routinely reported data, and patient/public input was not sought for study planning, implementation or dissemination. The study was focused on questions of importance to patients diagnosed with TB—what determines the success of treatment, and whether there are area-based differences in treatment outcomes that could be improved through changes in care practices. As a retrospective analysis of existing data, the study placed no burden on patients or public. All analyses were conducted using R.²⁶

Results

There were 356 119 individuals treated for TB over the 2015–2018 study period. We excluded patients previously diagnosed with TB (n=68 519), patients diagnosed with rifampicin resistance (n=2584), patients who had a change in regimen due to adverse event or identified drug resistance (n=2019), patients transferred between providers during therapy (n=20 306), patients diagnosed with TB post-mortem (n=2695), patients with missing treatment outcome values (n=10 786), patients with illogical combinations of exposure variables (n=56) and patients who could not be linked to a municipality (n=42). Records for 259 449 individuals were included in the analysis, representing individuals treated for TB for whom treatment outcomes could be categorised as one of treatment completion, loss to follow-up, death during treatment or regimen failure. Table 1 reports the distribution of individuals across patient-level exposure variables. Within this study cohort, 19.7% of individuals experienced an unsuccessful treatment outcome (death on treatment 7.8%, regimen failure 0.1% and loss to follow-up 11.9%).

Table 1. Baseline demographic information of the study population.

Variables	Number (%)	Variables	Number (%)
Age group		HIV
0–4	3018 (1.2%)	Positive	23 326 (9.0%)
5–14	4841 (1.9%)	Negative	191 087 (73.7%)
15–24	49 260 (19.0%)	Other	45 036 (17.4%)
25–34	58 818 (22.7%)	Smoking
35–44	47 672 (18.4%)	Yes	53 385 (20.6%)
45–54	39 908 (15.4%)	No	187 998 (72.5%)
55–64	30 693 (11.8%)	Other	18 066 (7.0%)
65–74	15 883 (6.1%)	Alcohol
75–84	7362 (2.8%)	Yes	41 714 (16.1%)
85+	1994 (0.8%)	No	202 648 (78.1%)
Sex		Other	15 087 (5.8%)
Male	177 304 (68.3%)	Drug
Female	82 145 (31.7%)	Yes	30 373 (11.7%)
Race		No	209 756 (80.8%)
White	82 410 (31.8%)	Other	19 320 (7.4%)
Black	31 402 (12.1%)	Incarcerated
Yellow	1825 (0.7%)	Yes	24 881 (9.6%)
Mixed	122 111 (47.1%)	No	218 659 (84.3%)
Indigenous	2908 (1.1%)	Other	15 909 (6.1%)
Other	18 793 (7.2%)	Homeless
Education		Yes	6170 (2.4%)
No education	11 259 (4.3%)	No	236 182 (91.0%)
Incomplete 1st-4th grade	29 382 (11.3%)	Other	17 097 (6.6%)
Complete 1st-4th grade	60 756 (23.4%)	Immigrants
Complete 5th-8th grade	47 907 (18.5%)	Yes	1543 (0.6%)
Complete high school	24 135 (9.3%)	No	236 944 (91.3%)
Any higher education	16 590 (6.4%)	Other	20 962 (8.1%)
Other	69 420 (26.8%)	Type of TB
Diabetes		Pulmonary	217 456 (83.8%)
Yes	19 934 (7.7%)	Extrapulmonary	7500 (2.9%)
No	223 788 (86.3%)	Both	34 493 (13.3%)
Other	15 727 (6.1%)

‘Other’ categories for variables HIV, diabetes, illicit use of drugs, alcohol, smoking, patients in vulnerable circumstances (incarcerated, homeless, immigrant) describe situations where the value was missing or could not otherwise be assigned to Yes or No.

Prediction model for patient-level factors

We assessed the performance of several candidate prediction models. The best performing model was a logistic regression model with random effects included for all two-way interactions of patient covariates, with a Brier score of 0.142 (table 2). This model was among the best-performing for each statistic assessed, and the calibration plot showed good calibration for different levels of predicted risk. When we fit this model to data for 2015–2017 and assessed performance in the 2018 data, the out-of-sample predictive accuracy was also good (online supplemental figure S1).

Table 2. Out-of-sample performance of candidate models used to adjust for patient characteristics.

Model	Brier score	C-statistic	Overall calibration	Calibration slope	Hosmer-Lemeshow statistic
Linear regression
Main effects only	0.143	0.708	−0.000	0.998	697.65
All two-way interactions	0.142	0.713	0.001	0.971	242.61
Logistic regression
Main effects only	0.143	0.708	−0.000	0.998	82.90
All two-way interactions	0.142	0.714	−0.000	0.964	40.85
All two-way interactions (random effects)*	0.142	0.715	−0.000	0.995	17.00
Boosted regression tree
Linear model
Main effects only	0.144	0.704	0.000	1.222	775.58
Interaction terms: two-way	0.143	0.709	−0.000	1.091	214.26
Interaction terms: three-way	0.142	0.711	0.000	1.038	114.15
Logistic model
Main effects only	0.144	0.703	0.000	1.230	970.38
Interaction terms: two-way	0.143	0.710	−0.000	1.110	318.51
Interaction terms: three-way	0.142	0.711	0.000	1.059	135.60
Ridge regression
Main effects and all two-way interactions	0.144	0.696	−0.000	1.011	92.89

The Brier score quantifies overall predictive performance, with lower values indicating better performance. The C-statistic measures the ability of the model to distinguish patients with different treatment outcomes, with higher values indicating better discrimination. The overall calibration indicates whether the model correctly estimates the overall fraction of the study population experiencing unsuccessful treatment, with a value of zero indicating perfect calibration. When paired with good overall calibration, a calibration slope close to 1 indicates that the model is well calibrated for high- and low-risk individuals. The Hosmer-Lemeshow statistic provides an additional measure of calibration within different risk levels. We estimated this statistic using 10 bins.

^*In this model (logistic regression, all two-way interactions with intercepts), a random effects distribution was applied to all two-way interactions between the covariates included in the model, to regularise these terms and reduce overfitting.

State-level outcomes

The excess risk of unsuccessful treatment was calculated by comparing treatment outcomes in each state to the national average. This overall excess risk was then decomposed into the component attributable to patient-level factors and the component attributable to area-level factors (table 3). Overall, Rio Grande do Sul had the highest proportion of patients experiencing unsuccessful treatment (7.0 (95% CI 6.4 to 7.7) percentage points excess risk compared with the national average), and Acre had the lowest (12.4 (95% CI 11.1 to 13.7) percentage points below the national average). Differences in treatment outcomes attributed to patient-level factors ranged from a 2.8 (95% CI 2.5 to 3.0) percentage points excess risk in Minas Gerais to a 4.3 (95% CI 3.9 to 4.8) percentage point lower risk for Acre. Differences in treatment outcomes attributed to area-level factors ranged from 5.2 (95% CI 4.6 to 5.9) percentage points excess risk in Rio Grande do Sul to an 8.1 (95% CI 6.8 to 9.4) percentage point lower risk for Acre. Across states, the variation in treatment outcomes due to area-level factors (SD 2.71) was greater than for patient-level factors (SD 1.68).

Table 3. Excess risk of unsuccessful treatment outcomes by state, decomposed into the contribution from patient-level and area-level factors, 2015–2018^*.

Name	Excess risk of unsuccessful treatment (percentage points)
	Total ($E^{All}$)	Due to patient-level factors ($E^{Patient}$)	Due to area-level factors ($E^{Area}$)
Rio Grande do Sul	7.0 (6.4, 7.7)	1.8 (1.6, 2.1)	5.2 (4.6, 5.9)
Mato Grosso do Sul	3.8 (2.4, 5.3)	−0.3 (−0.7, 0.2)	4.1 (2.7, 5.5)
Goiás	3.1 (1.7, 4.5)	2.3 (1.8, 2.8)	0.8 (−0.5, 2.1)
Rio de Janeiro	2.7 (2.3, 3.1)	0.4 (0.2, 0.5)	2.3 (1.9, 2.7)
Amazonas	2.7 (1.9, 3.4)	−0.1 (−0.3, 0.1)	2.7 (2.0, 3.5)
Mato Grosso	1.7 (0.4, 2.9)	0.7 (0.3, 1.1)	1.0 (−0.2, 2.2)
Minas Gerais	1.1 (0.3, 1.8)	2.8 (2.5, 3.0)	−1.7 (−2.4, –1.1)
Sergipe	0.7 (−0.8, 2.3)	0.2 (−0.3, 0.7)	0.5 (−1.0, 2.1)
Ceará	0.7 (0.0, 1.4)	1.0 (0.8, 1.3)	−0.4 (−1.0, 0.3)
Pernambuco	0.6 (0.0, 1.2)	1.8 (1.6, 2.0)	−1.2 (−1.8, –0.6)
Roraima	0.6 (−2.5, 3.7)	−1.1 (−2.0, –0.2)	1.7 (−1.3, 4.7)
Paraíba	0.5 (−0.9, 1.9)	0.4 (−0.1, 0.8)	0.1 (−1.2, 1.5)
Rondônia	−0.3 (−2.0, 1.3)	−1.9 (−2.4, –1.4)	1.5 (−0.1, 3.3)
Rio Grande do Norte	−0.4 (−1.7, 1.0)	1.2 (−0.8, 1.6)	−1.6 (−2.8, –0.3)
Alagoas	−0.6 (−2.0, 0.8)	0.9 (0.5, 1.3)	−1.5 (−2.8, 0.2)
Maranhão	−0.7 (−1.6, 0.2)	−1.6 (−1.9, –1.3)	0.9 (0.1, 1.8)
Santa Catarina	−0.9 (−1.8, 0.1)	−0.8 (−1.1, –0.5)	−0.1 (−1.0, 0.9)
Bahia	−1.3 (−1.9, –0.6)	1.6 (1.4, 1.8)	−2.9 (−3.5, –2.3)
Federal District	−1.5 (−3.8, 0.7)	−0.4 (−1.2, 0.4)	−1.1 (−3.3, 1.0)
Pará	−1.9 (−2.6, –1.3)	0.0 (−0.2, 0.2)	−1.9 (−2.6, –1.3)
São Paulo	−2.7 (−3.0, –2.4)	−1.7 (−1.8, –1.6)	−1.0 (−1.3, –0.8)
Tocantins	−3.0 (−5.8, 0.0)	−2.0 (−2.9, –1.0)	−1.0 (−3.8, 1.8)
Paraná	−3.1 (−3.9, –2.3)	−1.5 (−1.8, –1.3)	−1.6 (−2.4, –0.8)
Espirito Santo	−3.2 (−4.3,–2.1)	−1.0 (−1.4,–0.6)	−2.2 (−3.2,–1.1)
Amapá	−4.4 (−6.9,–1.8)	−3.2 (−4.0,–2.5)	−1.1 (−3.6, 1.4)
Piauí	−5.1 (−6.6,–3.6)	0.6 (0.1, 1.1)	−5.8 (−7.1,–4.3)
Acre	−12.4 (−13.7,–11.1)	−4.3 (−4.8,–3.9)	−8.1 (−9.4,–6.8)

^*$E^{All}$ represents the percentage point difference in the fraction of individuals experiencing unsuccessful treatment in each state compared with the national average, such that positive values indicate a higher percentage with unsuccessful treatment. $E^{Patient}$ represents the excess risk of unsuccessful treatment outcomes due to patient-level factors, and $E^{Area}$ represents the excess risk of unsuccessful treatment outcomes due to area-level factors, such that $E^{All}$ = $E^{Patient}$+ $E^{Area}$.

Municipality-level outcomes

We were able to estimate outcomes for 4911 of the 5572 municipalities in Brazil, with the remaining 661 municipalities (12%) excluded as they did not have any TB case notifications meeting the inclusion criteria during the study period. Figure 1 shows the municipality-level variation in treatment outcomes across Brazil, separated into the excess risk attributable to patient-level factors ($E^{Patient}$, figure 1A) and area-level factors ($E^{Area}$, figure 1B). For the excess risk attributable to patient-level factors, the highest risks were estimated for the municipalities of Goiânia (Goiás state, excess risk=5.4 (95% CI 4.3 to 6.5) percentage points), Belo Horizonte (Minas Gerais state, excess risk=5.3 (95% CI 4.6 to 6.0) percentage points) and Belford Roxo (Rio de Janeiro state, excess risk=5.2 (95% CI 4.4 to 6.0) percentage points). The lowest excess risks due to patient-level factors were estimated for the municipalities of Macapá (Amapá state, excess risk= −4.0 (95% CI −4.9 to –3.1) percentage points), Rio Branco (Acre state, excess risk= −4.0 (95% CI −4.6 to –3.3) percentage points) and São Vicente (São Paulo state, excess risk= −3.5 (95% CI −4.0 to –2.9) percentage points). For the excess risk attributable to area-level factors, the highest excess risks were estimated for the municipalities of Porto Alegre (Rio Grande do Sul state, excess risk=11.4 (95% CI 10.3 to 12.5) percentage points), Campo Grande (Mato Grosso do Sul state, excess risk=10.8 (95% CI 8.6 to 13.0) percentage points) and Nova Iguaçu (Rio de Janeiro state, excess risk=10.0 (95% CI 8.4 to 11.6) percentage points). The lowest excess risks were estimated for the municipalities of Rio Branco (Acre state, excess risk= −9.1 (95% CI –11.4 to –6.8) percentage points), Ribeirão Preto (São Paulo state, excess risk= −6.9 (95% CI −9.5 to –4.4) percentage points) and São José dos Campos (São Paulo state, excess risk= −6.6 (95% CI −9.3 to –3.9) percentage points). Figure 2 reports excess risk outcomes for major municipalities, including the state capitals plus the top 50 municipalities by total numbers of TB cases during the study period.

Figure 1. Excess risk of unsuccessful treatment attributable to patient-level and area-level factors, for Brazilian municipalities 2015–2018.* *(A) The distribution of municipalities by decile of excess risk of unsuccessful treatment attributable to patient-level factors ($E^{Patient}$). (B) The distribution of municipalities by decile of excess risk of unsuccessful treatment attributable to area-level factors ($E^{Area}$). Higher decile (darker shading) indicates a greater risk of unsuccessful treatment.

Figure 2. Excess risk of unsuccessful treatment attributable to patient-level and area-level factors, for major municipalities.* (A) The excess risk of unsuccessful treatment due to patient-level factors ($E^{Patient}$). (B) The excess risk of unsuccessful treatment outcomes due to area-level factors ($E^{Area}$). Figure shows results for the top 50 municipalities by total TB cases, plus state capitals. Horizontal bars indicate 95% uncertainty intervals. Letters in parentheses indicate the state for each municipality. TB, tuberculosis.

Online supplemental figure S2 shows the joint distribution of municipality-level estimates for excess risk due to patient-level factors and area-level factors, respectively. These two outcomes had a very weak negative correlation (rank correlation coefficient −0.031 (95% CI −0.059 to –0.003), p <0.05), indicating that in general, differences in patient-level factors do not explain municipality-level variation in treatment success rates. As with state-level outcomes, the municipal-level variation in treatment outcomes due to area-level factors (SD 4.27) was greater than for patient-level factors (SD 2.52).

Secondary outcome: excess risk of loss to follow-up

We re-estimated results for an alternative outcome defined as the excess risk of loss to follow-up. Online supplemental figure S3 shows the distribution of municipality-level estimates for the excess risk of loss to follow-up attributable to patient-level and area-level factors (online supplemental figure S3A,B, respectively). Municipality-level results for this secondary outcome were positively associated with the results of the main analysis, with a correlation coefficient of 0.56 (95% CI 0.55 to 0.58); p <0.05) for $E^{Patient}$ and a correlation coefficient of 0.77 (95% CI 0.75 to 0.78; p <0.05) for $E^{Area}$. For $E^{All}$, the correlation coefficient was 0.64 (95% CI 0.63 to 0.66; p <0.05).

Discussion

In this study, we assessed how the outcomes of TB treatment vary across Brazil according to patient-level and area-level factors, using data on 259 449 individuals initiating TB treatment over the 2015–2018 period. At both state and municipality level, we found that variation in the fraction of individuals experiencing an unsuccessful treatment outcome due to area-level factors—factors that cannot be explained by differences in patient characteristics—was substantially greater than the variation due to patient-level factors. For example, in our analysis, the difference between best and worst performing states according to patient-level factors represented a 7.1 percentage point increase in the fraction of individuals experiencing an unsuccessful treatment outcome. In contrast, the difference between best and worst performing states according to area-level factors represented a 13.3 percentage point increase, almost twice as great. A similar result was found at the municipality level. This variation due to area-level factors represents a major difference in the numbers of patients experiencing treatment success. For example, if all states were able to improve treatment outcomes to be at least as good as the top quartile of states, this would represent an average of 1142 additional TB patients achieving successful treatment every year in the study population, equivalent to 1.8% of total notifications.

By adjusting for differences in treatment outcomes due to patient-level factors, the residual variation (as quantified by $E^{Area}$) represents variation that cannot be attributed to differences in the individual risk factors recorded in the notifications data. While some of this variation may reflect the impact of individual risk factors not included in the notification record, it is likely that a major part of this variation represents the impact of differences in health system organisation and clinical practices across states and municipalities. In Brazil, TB policy and clinical guidelines are set at the national level, but how these policies are implemented varies across the health system. This variation is reflected in large inter-state differences in the fraction of patients treated via directly observed therapy and how care is distributed across primary, secondary and tertiary health system levels. Practically, such variation could produce differences in travel distances and other barriers that patients face to initiate and attend treatment, the range of support services provided during treatment or the level of effort devoted to following up on patients who have missed an appointment. It could also reflect differences in treatment access across geographic areas, with delayed treatment initiation as a possible cause of poor treatment outcomes. These geographic differences could be substantial, with notable differences in treatment coverage estimated at state and municipal level.^{13 14} Beyond healthcare access and quality, area-level variation in treatment outcomes could also include local socio-environmental factors that influence treatment adherence and completion. Such factors may include public initiatives to address social inequalities (eg, food vouchers, cash transfers) and social protection. Variation in outcomes due to patient-level factors (as quantified by $E^{Patient}$) has been investigated in prior studies,²¹,²⁴ with treatment mortality varying by age and HIV status, and loss to follow-up varying by factors tied to socioeconomic disadvantage (limited education and harmful health behaviours).

In addition to describing the variation in treatment outcomes across Brazil, this analysis identified specific states and municipalities that stood apart from other areas. For patient-level factors, Acre and Amapá states had low excess risks of unsuccessful treatment (>3 percentage points below the national average), while Goiás and Minas Gerais had high excess risks (>2 percentage points above the national average). While it is difficult for TB programmes to directly change the risk profile of individuals diagnosed with TB, understanding the differences between these states may reveal important differences in how TB risk is distributed across communities. For area-level factors, Acre and Piauí states had low excess risks of unsuccessful treatment (>5 percentage points below the national average), while Rio Grande do Sul and Mato Grosso do Sul had high excess risks (>4 percentage points above the national average). For the high-performing states, it may be that they have adopted care practices that allow them to achieve better outcomes. If true, these practices could potentially be employed in other areas to improve performance. Less optimistically, it is possible that these states have systematic problems with how treatment outcomes are being reported or have weaker case detection among the vulnerable populations most at risk of poor treatment outcomes, such that these individuals are never diagnosed and therefore not represented in treatment outcomes data. In both cases, further investigation would be valuable. Similarly, for those states for which this analysis estimated poor outcomes, there may be opportunities to improve treatment outcomes through further investigation to reveal the causes of poor treatment outcomes. A similar approach could be taken to studying municipalities that had results substantially higher or lower than the national average.

While previous studies have examined subnational differences in TB treatment outcomes, this is the first study that we are aware of that has systematically adjusted for differences in treatment outcomes due to patient level factors and decomposed overall rates of unsuccessful treatment to describe the contribution of both patient-level and area-level factors. We were able to do so by using a large and nationally representative dataset, allowing relatively precise inferences about how treatment outcomes vary across Brazil.

This analysis has several limitations. First, while the TB notification data used for this analysis include a range of relevant variables, it is possible that there are other individual-level factors that influence treatment outcomes that are not recorded in available data. If the prevalence of these factors differs systematically across states and municipalities, their omission from the analysis could bias the results estimated for individual geographic areas. Second, this analysis relies on the validity of treatment outcome reporting, and if there were systematic inconsistencies in how clinicians assign patients to treatment outcome categories, this would undermine the analysis. This is more likely to affect the secondary outcome (loss to follow-up), as a patient who dies while on treatment may be misclassified as lost to follow-up unless there are robust efforts to follow up on patients who stop attending clinic visits. Similarly, treatment outcomes were unavailable for some patients (treatment outcome recorded as transfer or missing, 9% of total sample) that would otherwise have been included in the analysis. If outcomes for these patients were systematically different from those included in the analysis, this could bias the results. Third, we did not attempt to identify the reasons why certain states and municipalities were achieving better or worse outcomes than others. This investigation represents a priority for future work and is a necessary step for translating this study’s findings into programme improvements. Fourth, this analysis attributed each patient’s treatment outcomes to their state and municipality of residence. While the approach we used will likely be unproblematic for state-level analysis, it is possible that some municipalities have a non-negligible fraction of patients travelling to neighbouring municipalities to receive care. Fifth, we did not consider the possibility of spatial dependencies in the regression models. This approach was taken so that the results estimated for individual states and municipalities would be independent of their neighbours. A consequence of this approach is that any spatial patterns may be underestimated.

In summary, our study identified substantial variation in TB treatment outcomes across states and municipalities in Brazil, which could not be explained by differences in patient-level factors. Further research to reveal the reasons for these differences is urgently needed to reduce geographic disparities in TB care across Brazil and increase the fraction of patients who successfully complete TB treatment.

Conclusion

Using prediction modelling, we adjusted for the case-mix of individuals treated for TB and estimated the variation in TB treatment outcomes across states and municipalities in Brazil. For both states and municipalities, we found greater variation in outcomes due to area-level factors compared to patient-level factors. This variation is likely a result of differences in health system structure, clinical practices and the role of socioeconomic factors. Further research is necessary to understand these differences and to shape effective strategies to improve the quality and effectiveness of TB care in Brazil.

Data availability statement

Data are available in a public, open access repository.