Impact of socioeconomic factors and health determinants on preterm birth in Brazil: a register-based study

In order to analyze the predictive relationships between PTB and health coverage determinants, we first aimed to understand PTB distribution in Brazil. We mapped the average spatial distributions of PTB rates (number of PTB/100 births) across two time periods, 2008–2012 and 2013–2017, to reduce possible nonspecific and irregular variations which can lead to misunderstandings of the results. We performed spatial autocorrelation to visualize clusters of PTB for both time periods across Brazil municipalities. Spatial autocorrelation measures the presence of dispersion or clustering in a region through Moran’s I (Index). The Moran's I index assumes a positive or negative value varying between -1 and 1, with a value of zero representing the hypothesis of spatial interdependence [37]. However, since a negative global Moran’s I value may not necessarily indicate the absence of spatial correlation on a local level [38], local indicators of spatial association (LISA) were applied by calculating and evaluating the significance of each city’s Moran’s I at alpha = 0.05, to find de clusters with high rates of PTB. The clusters are visualized on choropleth maps and characterized as follows: high-high (a city with a high PTB rate is close to others that also exhibit high PTB rates), high-low (a city with a high PTB rate is adjacent to others that have low PTB rates), low–high (a city with a low PTB rate is adjacent to others that have high PTB rates), low-low (cities with low PTB rates adjacent to others that also have low PTB rates) and NS (no significance, in which cities with high PTB rates are surrounded by cities with both high and low PTB rates).

We built a GWR model, aggregating data from socioeconomic status, structure of health services, and primary care work process seeking to forecast the PTB rate for each Brazilian municipality. To assess the spatial effect of social determinants on PTB rates, we performed an additional analytical step split into two approaches. The first analytical approach aimed to reduce the dimensionality of the predictors through principal component analysis (PCA) to meet the statistical requirements of GWR analysis. This first step generated factors summarizing the raw indicators selected for the analysis. These resulting factors were used as proxies of the predictors concerning socioeconomic status and the structure and work process of primary care services. The second analytical step used the GWR model to estimate the levels of PTB rates in the Brazilian municipalities. The predictors considered for GWR analysis came from the factors resulting from dimensionality reduction.

Since each domain (socioeconomic status, primary care work process, and structure of health services) consists of multiple variables, it is important to avoid using variables that are highly correlated due to multicollinearity issues. The use of several variables with overlapping variance can artificially inflate the portion of variance explained by each, thus creating a validity problem. The main strategy to overcome this is using dimensionality reduction approaches. Therefore, we used PCA on the indicators concerning socioeconomic status, structure of health services, and primary care work process to obtain a smaller, non-overlapping, number of factors representing each of the predictors. PCA is an orthogonal linear transformation [39] such that the largest eigenvalues (principal components) can be used to reconstruct a large amount of the variance of the original data [40]. The appropriate number of variables to retain for each domain was determined using component loadings and Scree plots (created to display total variation in each domain explained by each component). One factor was retained for each one of the three dimensions considered in the present study. The three factors were considered as predictors to generate a spatially weighted regression model aiming to forecast the PTB rate by municipality.

To understand associations between PTB and factors from the three health coverage domains, we adopted an approach based on spatially weighted regression models. The spatial dependency highlights the situations in which the local of the occurrence of a phenomenon can be at least partially explained by the geospatial perspective. To verify if the observed behavior regarding the PTB distribution over the Brazilian territory is modulated by the spatial dimension, we first ran a regression model using the Ordinary Least Square (OLS) approach without weighting the model for the spatial dimension. The OLS regression model minimizes the sum of square differences between predicted and observed values, leading to model parameter estimates without geographic representation [41].

The residuals of the regression model represent the magnitude of the distance between the forecasted values and the actual ones. Coefficients of the model must be statistically significant, without heteroscedasticity, and residuals normally distributed without spatial autocorrelation [42]. If the residuals of the OLS model are spatially correlated, this suggests the presence of spatial dependency of the object being studied. Thus, residuals from the initial OLS model were analyzed for spatial self-correlation using Moran’s I to evaluate the amount of PTB that could be explained by the spatial perspective. Once spatial dependency is observed, the best suitable model to forecast the variation in PTB rates is a GWR model. Since the residuals from the OLS predictive PTB analysis showed a strong spatial self-correlation, we used GWR to identify possible local associations and demonstrate the spatial effect of the multivariate model. GWR is a spatial analysis technique that takes non-stationary variables like environmental characteristics into account, to model predictive relationships [43].

GWR models use multivariate regression to evaluate associations between potential predictors and an outcome measure, weighting the results by the locality of occurrence [44]. The GWR model tested the PTB rate by municipality as the outcome measure, and the PCA factors regarding the socioeconomic status, structure of health services, and primary care work process as predictors. All analyses were performed using RStudio [45]. PCA was conducted using the psych package in R. The spatial self-correlation and OLS model were processed using GeoDa software 1.10.0.8 (Spatial Analysis Laboratory, Urbana, IL) [46], and the GWR model by GWR 4.0 [47]. Choropleth maps were created using QGIS 2.14 software [48].