Study design
The primary data source for this study is the cross-sectional 2013–2014 Demographic and Health Surveys for the DRC which is joined with remote-sensed environmental measures and land use data for mining and logging concessions extracted to DHS survey cluster locations. The DHS was administered using a multi-stage cluster survey design to represent the population of the DRC26. Briefly, survey clusters were selected to be representative of all 26 DRC provinces. Within clusters, households were randomly selected proportional to the population size, and within each household, adults ages 15–59 years were consented, interviewed, and asked to provide a dried blood spot (DBS) sample. Only adults who provided a DBS and consented for biospecimen use in future studies were included in this analysis. The outcome of prevalent malaria infections in the DRC was measured through PCR detection of the P. falciparum lactate dehydrogenase gene from DBS samples collected during DHS administration as described previously12.
The main exposures were residence within 15 km of a mining concession and residence within 15 km of a logging concession. Additional covariates included individual-level variables for participant age, sex, use of a long-lasting insecticidal net (LLIN), education, and occupation; household variables for wealth, house roofing material, and the ratio of the number of household members using a bed-net to the total number of household members; and cluster variables for elevation, temperature, precipitation, vegetation, percentage of land cover identified as cropland, grassland, forest, and flooded/swamp land. All individual and household variables were obtained through the DHS. Occupation was recoded such that the manual labor and army category included laborers in mining and logging industries. Cluster variables were extracted from various satellite imagery platforms and other spatial datasets; the methods are described in more detail in the “Appendix”. The main exposures were extracted from geographic data sources as described below.
Mining and logging concession data were obtained from the Global Forest Watch online repository27. Mining concessions were subset to only include operations that were active or in remediation spanning the DHS study years (2013–2014); logging concessions only included active operations during 2013. Distance to a mining or logging concession was measured from each cluster location to the boundary of a concession. Clusters were considered exposed to mining or logging if they were located within 15 km of a concession. This distance was chosen to account for the estimated 10 km maximum flight distance of a blood-fed mosquito5, with an additional 5 km to compensate for boundaries and non-residential land near the concessions. This range also accounts for the 5–10 km random spatial offset implemented by the DHS. Locations of mining and logging concessions along with cluster locations were mapped across the DRC. All maps were created in ArcGIS version 10.7.1, shapefiles for administrative boundaries were obtained from GADM.org.
Data analysis
Characteristics of the study population were evaluated across quantiles of P. falciparum cluster prevalence and grouped by individual, household, and cluster level variables. To further examine distributions of malaria interventions and risk factors such as age, sex, LLIN use, occupation, household wealth, and household roof materials by mining and logging exposure, we compared mining exposed and logging exposed clusters with mining and logging doubly unexposed clusters stratified by urban and rural residence.
We then modelled the prevalence odds of malaria across the DRC using hierarchical logistic regression models to account for the nested structure of the DHS data and to allow for inclusion of spatially varying effects. Models were implemented in a Bayesian framework using Integrated Nested Laplace Approximation (INLA) and stochastic partial differential equations for spatial effects28. In all models, we included two separate indicator terms for proximity to a mining concession and to a logging concession; since these areas are non-overlapping, the referent condition for each of these exposures is therefore locations exposed neither to mining nor to logging.
The model fitting process followed two approaches. The first approach evaluated population-level effects of mining and logging on malaria prevalence adjusting for covariates and accounting for cluster-level random effects, which were assumed to vary independently across clusters. The second approach retained covariates and the cluster-level random intercept from the first model and additionally incorporated a spatial field to account for confounding due to space. For the spatial approach, two models were constructed. The first included a spatially varying intercept which borrowed information from neighboring cluster locations assuming a Gaussian random field. The second spatial model explored possible residual confounding due to environmental covariates by allowing spatially varying slopes for temperature, precipitation, vegetation, elevation, and land cover classes while including both independently and spatially varying intercepts across clusters. We introduced spatially varying slopes to account for the unobserved vector population across the DRC. Temperature, precipitation, vegetation, elevation, and various land cover classes have been shown to influence vector composition, survival, and competence for P. falciparum5,23,25, and associations with these covariates may vary due to their effects on the unobserved vector population. Using the spatial modelling approach, we also constructed a smoothed predicted prevalence map of malaria across the DRC, additional details are in the “Appendix”.
For all models, confounding variables were selected based on a directed acyclic graph analysis and retained for adjustment if the 95% uncertainty interval (UI) of the variable excluded the null. Variables were coded as they were presented in the DHS with the exception of collapsing wealth into moderate or higher versus low wealth and recategorization of occupation as: professional, sales, or services; not working; manual labor or army; and agricultural work. All environmental variables were coded as continuous and scaled. Land cover variables were coded in intervals of 10 percentage points. Model comparison was done using Deviance Information Criterion (DIC), with the best fitting model having the smallest DIC29. All models were run using the ‘INLA’ package in R version 4.0.428, additional details are described in the “Appendix”.
Differences in urban and rural residence were considered an important potential source of bias. Urban residence has been associated with lower prevalence of malaria due to many factors including different vector habitats, better access to healthcare, improved housing construction, and overall higher wealth4,12. To address possible bias introduced by urban residence, we stratified all models by urban and rural residence based on the DHS classification of clusters as urban or rural.
A discrete set of confounding variables was identified from fixed effect models for mining and logging in rural and urban areas. The final adjustment set included age, sex, LLIN use, household wealth, temperature, precipitation, vegetation, and elevation. These variables had statistical or substantive significance and were adjusted for in all consecutive analyses.
Ethical approval for this study was obtained from the University of North Carolina Institutional Review Board (UNC IRB# 20-3175) and the Kinshasa School of Public Health. Informed consent was obtained from all participants and all methods were conducted in accordance with guidelines and regulations set forth by the UNC IRB and the Kinshasa School of Public Health.
Source: Ecology - nature.com
