Proximity to mining industry and respiratory diseases in children in a community in Northern Chile: A cross-sectional study

Background In a community in northern Chile, explosive procedures are used by two local industrial mines (gold, copper). We hypothesized that the prevalence of asthma and rhinoconjunctivitis in the community may be associated with air pollution emissions generated by the mines. Methods A cross-sectional study of 288 children (aged 6–15 years) was conducted in a community in northern Chile using a validated questionnaire in 2009. The proximity between each child’s place of residence and the mines was assessed as indicator of exposure to mining related air pollutants. Logistic regression, semiparametric models and spatial Bayesian models with a parametric form for distance were used to calculate odds ratios and 95 % confidence intervals. Results The prevalence of asthma and rhinoconjunctivitis was 24 and 34 %, respectively. For rhinoconjunctivitis, the odds ratio for average distance between both mines and child’s residence was 1.72 (95 % confidence interval 1.00, 3.04). The spatial Bayesian models suggested a considerable increase in the risk for respiratory diseases closer to the mines, and only beyond a minimum distance of more than 1800 m the health impact was considered to be negligible. Conclusion The findings indicate that air pollution emissions related to industrial gold or copper mines mainly occurring in rural Chilean communities might increase the risk of respiratory diseases in children. Electronic supplementary material The online version of this article (doi:10.1186/s12940-016-0149-5) contains supplementary material, which is available to authorized users.


Gold mine
Copper mine

C Multiple imputation process
Given the presence of missing data (Table 1), seven data set were generated and the Rubin's rules [1] were used to calculate the parameter estimates. Results of the multiple imputation process are given in Figure C.1. Figure shows the unadjusted Odds Ratios (OR) for each covariate after the imputation process, these were calculated using logistic regression models.

D Unadjusted STAR models
The models estimated were as follows: The resulting estimation are presented in Figures  Only for subjects who live within distance range from the mine. Shaded area is 95% Bayesian confidence interval.

Gold (km)
log(OR) Only for subjects who live within distance range from the mine. Shaded area is 95% Bayesian confidence interval. The random term S(s) represents a residual spatial component in the models and is independent from the presence of the mines. It reflects all the variation not accounted by the covariates and the distance function. Different approaches are used to model this term, depending the context, perhaps, Gaussian random field in the analysis of lattice data [2]. However, [3,4] and [5,6] (in the context of modelling of the association between risk and relation of putative source) used an approach with a thin-plate splines interpolate, this last approach was used in this paper to model this effect, this semi-parametric term is favourable because of the computational advantages.
For the spatial compound specification, given a set of T spatial nodes of the N children locations, S(s) has a low-rank representation following the form: with b is a T-dimensional vector of random coefficients to control of the spatial smoothing, Z(s) is the sth row of the design matrix In 3, Z T and Ω T are the spatial correlation matrix between the N children residential locations and the T nodes and that among the nodes, respectively. Both matrices are based in a isotropic spatial correlation function in the radial basis function, it means we used C(r) = r 2 log(r) where r is the Euclidean distance, resulting in Z T = [C{d(s, t)}] and Ω T = [C{d(t, t ′ )}], with t, t ′ = 1, . . . , t T , completing the thin-spline representation necessary in 2.
Finally, priors distributions for the thin plate approach in Equation 2 are as follows:

F.1 Prior elections
Election of the prior distributions have been study previously [2,5], suggesting special care in the priors' elections specially for α k and ϕ k . We followed the proposal in Li et al. [8], doing a one-to-one transformation of the parameters (u k , v k ) = (log(1 + α k ), log(ϕ k )).
We considered mutually independent prior Normal distributions on (u k , v k ), For α k and ϕ k . A simple transformation allows to calculate µ u k for a proposed value of the mean of α k , µ α , which is obtained from the STAR models. The hyperparameters µ u k and σ 2 u k for α k specify a distribution reflecting the point estimate and uncertainty obtained with the STAR models at the nearest distance. For ϕ k , µ v k and σ 2 v k reflected the decrease of the risk over distances from the edge of the plateau at which the risk had decreased.(33). The parameter δ k is distributed as a Gamma(κ 1 , κ 2 ). The hyperparameters κ 1 and κ 2 defined an informative prior reflecting the radius of the plateau that no observation at distances below than 0.87 km. The spatial effect S(s) collected the residual variation across the region not accounted for the potential confounders or by the proximity to the mines. S(s) was estimated using Bayesian thin-plate splines.
F.2 Sensitivity analysis for the parameter α As sensitivity analysis of the prior specification on the parameter α, posterior densities were obtained and compared from different choices of prior distributions for the Model 3 using asthma as outcome. The priors were defined as: (2) Models are conditioned to the election of a good prior distribution when the election is a uniform distribution, however the election of normal distribution on parameters transformation should be showed more stability. The election of prior flat or non informative distribution must be considerer to no use in the uniform case because the modes are sensitive to the election of the upper limit, it was also mentioned by other authors [2,7,8]. The parametrization proposed by Li et al. showed more robustness in our study. Posterior densities using others outcomes presented a similar behaviour (not shown).