Study design, population, and ethical issues
This study has a quasi-experimental cohort study design, based on the longitudinal information of 54.5 million individuals from January 1, 2004 to December 31, 2015 (the period for which tuberculosis data were available). First, we constructed a conceptual framework to explain the mechanisms of possible effects of CCT on TB outcomes and to drive the analysis (Figure 1).15 The study population was achieved by selecting a subgroup of individuals of the 100 Million Brazilians Cohort16, a consolidated cohort created through the validated linkage17 between the Federal Government Unified Registry for Social Programs (Cadastro Único) – that gathers data from the poorest half of the Brazilian population, identifying and characterising low-income families for social programs eligibility, and including information on exposure to the BFP - and health-related datasets from the Brazilian Ministry of Health's (Appendix, p.3).
This study was approved by the Research Ethics Committee of the Institute of Collective Health of the Federal University of Bahia (ISC/UFBA), under number 41691315.0.0000.5030 (Assessment nº:3.783.920).
Data sources, outcomes, and intervention
Two individual-level health-related datasets were linked to Cadastro Único (CADU): the Notifiable Diseases Information System (SINAN) and the Mortality Information System (SIM). SINAN contains records of notifiable diseases, including TB. SIM registers deaths by all causes, according to International Classification of Diseases (ICD-10). The linkage codes and algorithms were built based on five identifiers: date of birth, municipality of residence, sex, name, and mother`s name of the individual in each database. The CADU and the health information datasets (SIM and SINAN) were individually matched in two steps, using the CIDACS-Record Linkage tool (Appendix, p.3). The quality of each link between CADU, SINAN, and SIM has been extensively evaluated and validated.17 An aggregate-level longitudinal dataset - containing a wide range of yearly municipal-level information on TB endemicity levels, municipal infrastructures, and healthcare resources - was also linked to the cohort through the individuals` municipal code of residence.
Tuberculosis outcomes defined for the study were: incidence, mortality, and case-fatality rates. The beneficiary group was defined as eligible individuals who received BFP benefits, and their exposure started with receipt of the benefit, until the end of their follow-up. The non-beneficiary group was defined as individuals who had never benefited from BFP throughout their follow-up period. In case of non-receipt of the benefits, eligible individuals were classified in the non-beneficiary group (Appendix p.4).
Statistical Analyses
First, in the descriptive analysis, we estimated the rates of the study outcomes as follows: i) TB incidence: new TB diagnoses divided by person-years at risk and multiplied by 100,000; ii) TB mortality: TB deaths, divided by person-years at risk and multiplied by 100,000; and iii) case-fatality rate: TB deaths among people affected by TB, divided by person-years at risk and multiplied by 100. The follow-up time for each individual in the cohort, i.e., person-years, started on the date of entry into the cohort until the date of TB diagnosis (for TB incidence), the date of death due to TB (for TB mortality rate), the date of death from other causes, or the end date of the cohort (December 31, 2015). For TB case-fatality rate, the start date began with the date of diagnosis and ended with the TB-related death, the date of death from other causes, or the final date of the cohort. Afterwards, we performed a descriptive analysis of new people affected by TB and deaths according to each independent variable. At the individual level, the demographic and socioeconomic covariables were age, sex, self-identified race/ethnicity (white, Indigenous, Black and pardo - these last categories were analysed together), education, per capita expenditure (as a proxy for the per capita wealth and calculated as a percentage of the yearly minimum wage, categorised by tertiles), and year of entry into the cohort. At the family level, the independent variables were related to household characteristics: number of people, water supply, construction material, sanitation, garbage disposal, and lighting. At the municipal-level, the covariables were unemployment rate, Gini Index, and a set of variables related to health services: Family Health Strategy coverage (the main model of Primary Health Care in Brazil), number of doctors, nurses, and specialised clinics per 1,000 inhabitants. To control for any potential selection bias associating PBF implementation with endemic TB levels in the community, the mean TB incidence rate in the cohort during the study period was included as a covariate in the models. When the study outcome was the case-fatality rate, we also included clinical classification of TB, percent of directly observed therapy (DOT), AIDS comorbidity, and diabetes as independent variables. All the variables used in the study are described in the conceptual model (Figure 1).
To estimate the effect of BFP exposure on TB incidence, mortality, and case-fatality rates we used multivariable Poisson regression models, adjusted for all the relevant demographic and socioeconomic confounding variables listed above, with follow-up time as an offset variable, robust standard errors, and observations weighted through stabilised, truncated, inverse probability of treatment weighting (IPTW). Poisson regression models are common for cohort data analyses,18 and IPTW Poisson regression models have been used in quasi-experimental cohort studies which investigate the impacts of public and social policies on health outcomes, including several evaluation studies that used the 100 Million Brazilian Cohort.8–10,19 Moreover, in order to understand BFP effects heterogeneity, we fitted these IPTW Poisson regression models stratified by age, sex, race/ethnicity, education, and wealth- tertiles (per capita expenditure).
To confirm the robustness of the findings, we applied several sensitivity analyses (for details see Appendix, p.9-12): i) we fitted models with only individual-level variables and tested the inclusion of different aggregate-level variables, ii) we fitted the same regressions without the TB endemicity level variable, iii) we estimated and compared all models without IPTW, iv) to evaluate the adoption of per capita expenses as a proxy for wealth, we carried out the same analyses with other proxies, such as per capita income, v) we adjusted the same models with different specifications (including different sets of individual-level covariates, inclusion or exclusion of robust standard errors, only in municipalities with adequate vital information). Finally, to have a greater degree of confidence in the causal inference of our impact evaluation, we performed two different triangulation analyses,20 verifying the existence of BFP effects also using alternative methods: survival analysis with Cox multivariate regression and propensity score matching (PSM) (Appendix, p.13-14).
Role of the funding source
The funding source had no role in study design, data collection, data analysis, data interpretation, or the writing of the report.