2.2. Availability of data and materials
The analytical units in this study were the census tracts, based on data from the latest Population Census by the Brazilian Institute of Geography and Statistics (IBGE) in 2010. Secondary data were collected from three information sources for 2015 and 2016: (i) new cases of ZIKV residing in the city of Rio de Janeiro, reported to the Information System on Diseases of Notification (SINAN); (ii) sociodemographic data from the census tracts collected by IBGE (21); and (iii) data on addresses, rainfall, and FHS coverage for the years 2015 and 2016 from Pereira Passos Institute of the Rio de Janeiro Municipal Government (22).
Based on the SINAN database, new ZIKV cases were geocoded by residential address on the notifications, using API (Application Programming Interface) from Google Maps (23). Geocoded ZIKV cases were then aggregated by census tracts, allowing the construction of crude ZIKV incidence rates per 1,000 inhabitants. To control the incidence rates’ instability effect we used local empirical Bayesian smoothing, weighting the incidence rates of neighboring tracts (24). We then applied the rates’ logarithmic transformation with Bayesian smoothing to approach them to a normal distribution, thereby establishing the study’s outcome, hereinafter the “ZIKV incidence rate”.
Rainfall data were collected from 33 precipitation stations in the city of Rio de Janeiro from November 2015 to February 2016. These data were used to build the variable mean rainfall by census tract via spatial interpolation. Since this information was not measured in all the census tracts, but in 33 precipitation stations in the city of Rio de Janeiro, geostatistical techniques were used to estimate rainfall in areas without precipitation stations. This was used to create a continuous map with the estimated rainfall values for entire city of Rio de Janeiro through simple kriging. This method was chosen since it was expected that the mean monthly rainfall in millimeters from November 2015 to February 2016 would be constant across the surface. Next, the census tract map was superimposed on this map, which allowed calculating the mean rainfall for each tract in the selected period. To perform the spatial interpolation, a 5,000,000 × 5,000,000 grid was defined, since some areas were very small. After spatial rainfall analysis, the model with the best fit was spherical anisotropy effect (25).
Data on coverage by the Family Health Strategy (FHS) defined whether or not the census tract was covered by the FHS.
These databases were then merged, using the census tract number as the identifier field. This allowed a final database with ZIKV incidence rate per census tract as the dependent variable (outcome) and sociodemographic data, mean rainfall, and FHS coverage in each census tract as the independent variables.
2.3. Data analysis
We assessed the presence of spatial autocorrelation in the ZIKV incidence rates per census tract, using global Moran’s index. The Local Indicator of Spatial Association (LISA) was used to investigate spatial association patterns at the local level, disaggregating the global Moran’s index. LISA classifies the tracts based on the neighborhood matrix, in four groups: high/high, low/low, high/low, low/high. The first two groups correspond to positive associations between a census tract’s incidence rate and the respective neighbors’ rates, while the other two groups represent negative associations (26).
The theoretical model to represent the relations between the ZIKV incidence rate and possible associated factors was built from a combination of two conceptual models. “Model I” was proposed by Diderichsen, Augusto, and Perez (27). The authors summarized a diagram containing possible mechanisms of social inequality in relation to ZIKV infection. These mechanisms were: income, household conditions, vector density, vulnerability, susceptibility, social context, and health policies. “Model II” related the occurrence of natural disasters (earthquakes, floods) to the increase in the mosquito population and consequently to the rise in ZIKV incidence (28). According to the available data, this study proposed a new model including socioeconomic dimensions, household conditions, and health policies (model I) and rainfall (model II).
Figure 2 presents this study’s proposed model as a causal diagram (29) associated with the Structural Equation Model (SEM) (30). In this diagram, the ellipses correspond to the construct with two latent variables, i.e., which were not measured (socioeconomic status and household condition). The first latent construct has five indicators represented by the observed variables: INCOME < 1 MW (proportion of households with income less than 1 minimum wage); INCOME1to2MW (proportion of households with income from 1 to 2 minimum wages); BROWN_BLACK as the proportion of households with individuals self-identified as brown or black; LIVE_ALONE (proportion of households with persons living alone); NO_SCHOOL (proportion of illiterates). The latent variable “household conditions” is represented by 3 indicators: WATER_SYST (proportion of households with running water supply); SEWAGE_SYST (proportion of households connected to the public sewage disposal system); GARB_COLLECT (proportion of households with public garbage collection). The variable RAINFALL represents the mean monthly rainfall in millimeters from November 2015 to February 2016; TEAM_FHS shows whether the census tract was covered by Family Health Strategy teams.
One can thus hypothesize that census tracts with heavy social inequality (low income, no schooling, and no family support) can present worse living conditions, also impacting household conditions. Meanwhile, heavier rainfall in a given region an also affect the household, for example, with accumulation of water in mosquito breeding sites. These combined causes may have contributed to the increase in ZIKV cases. In addition, the effect of health policies such as the Family Health Strategy can promote educational activities in the fight against Aedes aegypti, which would tend to reduce the number of cases, while also offering greater access to health services, allowing a reduction in underreporting of ZIKV cases.
The analysis excluded 332 census tracts that represented areas with lagoons, forests, green areas, and individuals in institutionalized sites or without population information from the IBGE.
After confirmation of spatial autocorrelation, we assessed the relationship between the outcome (ZIKV incidence rate) and the independent variables through Spearman correlation analysis (ρ). Variables with ρ > 0.3 were tested in the model with interaction. The Spearman correlation matrix showed a statistically significant correlation with the study variables. However, there was strong correlation (indicating possible collinearity) between INCOME < 1 MW and BROWN_BLACK; INCOME < 1 MW and NO_SCHOOL; and BROWN_BLACK and NO_SCHOOL.
The model’s fit was assessed with Ordinary Least Squares (OLS), Spatial Lag Model (SAR), and Spatial Error Model (CAR). The first method includes the traditional linear regression approach, the second incorporates spatial dependence into the dependent variable, and the third method includes the spatial effect jointly in the model’s random component (error) (31). To choose the model with the best fit, we opted to compare the models according to the highest log-likelihood value and the lowest Akaike information criterion (AIC) value.
Diagnosis of collinearity was performed with the Variance Inflation Factor (VIF) with tolerance values less than 10 (32).
All the models’ residuals were assessed with the Moran index to quantify the degree of spatial dependence.
All the analyses were performed in the R statistical package, version 3.4.3 (33) and in GeoDa (34).