This retrospective cohort study uses a subsample of data from the 100 Million Brazilian Cohort,15 a consolidated cohort that used validated linkage algorithms16 to merge the Unified Registry of Brazilians (CadÚnico), a national tool used to determine eligibility for social protection programs, with the Brazilian National Surveillance System and the Brazilian Mortality Information System over a 9-year period. The Center for Data and Knowledge Integration for Health (CIDACS) developed this linkage (Supplementary Appendix).15,16 The study protocol17 was approved by the Research Ethics Committee of the Collective Health Institute at the Federal University of Bahia, number 41691315.0.0000.5030 (Assessment nº:3.783.920).
Dataset and Study Population
We selected a sample of 28.3 million Brazilian people aged ≥13 years, and registered between 2007 and 2015 from the 100 Million Brazilian Cohort. Individuals diagnosed with AIDS, or those already deceased before the start of the cohort follow-up (January 1, 2007), were excluded (description of the selection process in the Supplementary Figure S1 in appendix, p. 4). For the intersectional analyses, three markers of social differences - race, education and wealth – were selected based on their relevance for HIV/AIDS outcomes,8-10,18 including their strong health effects observed in a previous study developed with the same cohort18, and other preliminary analyses (Supplementary Table S1 in appendix, pp. 5-6). Intersectional analyses were conducted using a binary classification of the social difference markers: race (Whites/Asian or Blacks), education (higher or lower educational attainment) and wealth (higher or lower wealth). Other race categories (i.e., Brown and Indigenous) and the intermediate level of wealth were not presented in the main study results but were included in the estimated models (Supplementary Table S2 in appendix, pp. 8-10).
Variables and measures
Follow-up was calculated from the cohort entry date or AIDS diagnosis (case-fatality rates) until the time of censoring, which was determined by: a) AIDS diagnosis (incidence); b) AIDS-related death (mortality and case-fatality rates); c) death by other causes; or d) study end date (December 31, 2015) (Supplementary Chart S2 in appendix, p. 18). The criteria used to define AIDS cases were adapted CDC, the Rio de Janeiro/Caracas, and AIDS deaths. AIDS deaths were defined as the underlying cause of the ICD-10 codes B20 to B24 (Supplementary Charts S1 and S2 in appendix, pp. 16-18).
Associations between race, education, wealth, and AIDS-related outcomes were examined both individually and intersectionally. Thus, dummy variables were constructed with two- and three-way interaction terms, and categorized into: a reference group (Whites, higher educational attainment and wealth), groups with the presence of one social marker (isolated markers of social difference), and groups with an intersection of two or more social markers (combined markers of social difference) (Supplementary Figure S2 in appendix, p. 20). The following comparisons were thus analyzed: 1. Intersection between wealth and education: individuals with higher wealth & higher education (reference group) were compared with individuals of: (a) higher wealth & lower education, (b) lower wealth & higher education (isolated markers of social difference), and (c) lower wealth & lower education (combined markers of social difference); 2. Intersection between race and education: White/Asian individuals with higher education (reference group) were compared with individuals of: (a) White/Asian race & lower education, (b) Black race & higher education (isolated markers of social difference), and (c) Black race & lower education (combined markers of social difference); 3. Intersection between race and wealth: White/Asian individuals with higher wealth (reference group) were compared with individuals of: (a) White/Asian race & lower wealth, (b) Black race & higher wealth (isolated markers of social difference), and (c) Black race & lower wealth (combined markers of social difference); 4. Intersection between race, education, and wealth: White/Asian individuals with higher wealth & higher education (reference group) were compared with individuals of: (a) White/Asian race & higher wealth & lower education, (b) White/Asian race & lower wealth & higher education, (c) Black race & higher wealth & higher education (isolated markers of social difference), (d) White/Asian race & lower wealth & lower education, (e) Black race & higher wealth & lower education, (f) Black race & lower wealth & higher education, and (g) Black race & lower wealth & lower education (combined markers of social difference) (Supplementary Figure S2 in appendix, p. 20).
Gender was taken as an effect-modifier of the association between race, wealth, education, and AIDS outcomes, and was used in the stratification analysis. The following individual-, household-, and municipal-level variables associated with both the independent variables and outcomes were included in the multivariate analyses: geographic factors (Brazilian geographic region and area of residence); time since last received conditional cash transfers (Bolsa Família), household conditions, age, year of admission into the cohort; and health infrastructure (primary health care coverage, number of specialized clinics, physicians, and hospital beds per 1,000 inhabitants), average AIDS outcome rate (in the follow-up period) under study, and socio-environmental conditions (Gini index, extreme poverty and unemployment rates) in municipality of residence (Figure 1).
Statistical analysis
Descriptive statistics explored the distribution of the three-way intersection groups by the socioeconomic covariates. To estimate the crude (RR) and adjusted (aRR) rate ratios, multivariable Poisson regression models were used,19,20 using robust standard errors, clustered in the municipality of residence, and separately for each outcome. Poisson regression models that incorporate follow-up time as an offset variable to generate interpretable risk measures21 are common in cohort data analyses22 and have widely been used in similar observational studies using the 100 Million Brazilians Cohort.18-20
The intersections of markers of social differences were also evaluated through a wide range of interactions indicators on additive and multiplicative scales.5,23 This innovative approach allows to estimate the intersectionality in a more robust and comprehensive way, because it assigns a direction (positive, null, or negative) and a magnitude (assessing how much higher is the risk attributed to the combined characteristics) to the interaction under study, which is particularly useful for the implementation of more equitable public policies.24
We calculated several interaction measures: (a) Excess Risk (ER), that is, how much the risk of the presence of interaction exceeds the risk of none of the characteristics present;23 (b) Relative Excess Risk due to Interaction (RERI), that is, how much the excess risk due to the presence of interaction deviates from the null value (RERI>0 signals interaction);23,25 (c) Attributable Proportion (AP), which shows the proportion of cases due to the interaction of both exposures (AP>0 signals interaction);23,25 (d) Synergy Index (S), which reflects the direction of the interaction in relation to the null value (S=1 signals no interaction, S>1 synergy, and S<1 antagonism);23 (e) Multiplicative scale interaction measure for hazard ratios (MIM), which assesses the extent to which the effect of the interaction between exposures exceeds the product of the effects of the exposures seen separately (MIM>1 signals positive interaction, MIM<1 negative, MIM=1 null);24 and (f) Regression product terms, i.e. the RR estimate and the 95% confidence intervals (CI) obtained for the product term.24 Moreover, we evaluated at the same time three social markers, namely race, education, and wealth, which is innovative in intersectional research and even in HIV/AIDS research, and it has been possible because of our large data samples and our adaptations in the calculation of the indicators interaction (Supplementary Chart S3 in appendix, pp.22-23).
Adjusted Rate Ratios were calculated while controlling for all covariates (Figure 1). We implemented a number of strategies to assess the robustness of our findings (Supplementary Appendix, pp. 29-48): 1) Sensitivity analyses: alternative regression models were run (negative binomial regression; survival models, and multilevel models), to examine potential changes in the estimates based on model form specification; 2) Further adjustments for covariates: models were re-estimated with and without municipal-level variables; and 3) Data quality influence analyses: to determine the influence of municipal differences in surveillance data quality, we estimated models only for individuals living in municipalities with an adequate quality of vital information (QVI), according to validated criteria, and compared with the findings from all individuals under study. All analyses were performed using Stata® 15.0.
Role of the funding source
This study was funded by the National Institute of Allergy and Infectious Diseases - NAIDS/NIH, Grant Number: 1R01AI152938. The study funders played no role in the study design, data collection, data analysis, data interpretation, or writing of the paper. None of the authors were precluded from accessing study data, and they accepted the responsibility to submit the paper for publication.