Bayesian geostatistical variable selection and prediction of tropical diseases

A global commitment from governmental, non-profit, research and even profit organizations to combat tropical diseases has led to an increase of funding for implementing control interventions. Guidelines for controlling a disease commonly depend on its prevalence or incidence. Information on the disease risk distribution is important for successful control implementation. Spatial statistical modelling provides a framework to predict disease risk at high spatial resolution, assess disease dynamics and evaluate the effects of interventions.
In sub-Saharan Africa, estimates of soil-transmitted helminthiasis risk and of treatment requirements are lacking, mainly due to scarcity of georeferenced data and inaccessibility of the available ones. There is a need to bridge this gap for cost-effective disease control, monitoring and evaluation. Soil-transmitted helminthiasis is a poverty-related disease. Socioeconomic proxies (SES), such as socioeconomic status, access to safe water and sanitation (WASH) facilities, could improve predictive risk modelling. Socioeconomic data are available from household surveys and are georeferenced at village-level. It is unclear whether village-aggregated SES can improve predictions of disease risk. Brazil is one of the most affected countries with leishmaniasis. Despite the implementation of a notifiable system in the country, geostatistical analyses of leishmaniasis incidence are limited to few districts and provinces in the country. A countrywide analysis estimating the geographical and temporal distribution of the disease has not been carried out. There is a lot of progress in malaria control over the last years. Interventions are widely administered and repeated national surveys in Africa are conducted collecting spatial data on the disease risk and on a number of intervention coverage indicators. However, the effects of the ongoing interventions on malaria risk have not been analysed. Model formulations that can estimate the effects of malaria interventions in space and time have not been established.
This thesis aims to address the above gaps of knowledge by developing datadriven Bayesian geostatistical models. The specific objectives of this research are to: (i) assess the spatiotemporal distribution, identify risk factors and calculate the number of infected Brazilians with leishmaniasis (Chapter 2); (ii) predict the distribution of soil-transmitted helminth (STH) infection risk in sub-Saharan Africa, evaluate temporal trends, and provide spatiotemporally explicit estimates of people infected and of treatment requirements by country (Chapter 3); (iii) predict the distribution of STH risk in Cambodia and evaluate the predictive ability of SES proxies in geostatistical disease modelling using individual and village-specific SES (Chapter 4); (iv) provide geostatistical models with spatially varying covariate effects and assess variable selection formulations to estimate effects of malaria interventions on disease risk (Chapter 5); and (v) develop geostatistical models with spatiotemporally varying covariate effects and evaluate sensitivity of predictive process approximation for variable selection of large data (Chapter 6).
In Chapter 2, we apply Bayesian geostatistical negative binomial models to analyze reported incidence data of cutaneous and visceral leishmaniasis in Brazil covering a 10-year period (2001-2010). Particular emphasis is placed on estimating spatial and temporal patterns. The number of cases are predicted at province and country levels.
In Chapter 3, we analyze soil-transmitted helminth infection risk in sub-Sahara Africa. Data are obtained from a systematic review and analyzed using geostatistical models. Areas where data are lacking but a high infection risk is predicted are highlighted. We calculate anthelmintic treatment needs by country using World Health Organization guidelines.
Chapter 4 presents a geostatistical analysis of soil-transmitted helminth infections in Cambodia. The study pursues an in-depth investigation of the use of socioeconomic predictors in mapping poverty-related diseases. Additional to the country-level analysis with SES aggreagated at village level, separate analyses are carried out using individual-level SES proxies to assess and quantify their associations with soil-transmitted helminth infections. Analyses using individual and village-specific proxies are compared.
In Chapter 5, we provide geostatistical models with spatially varying coefficients for estimating effects of malaria interventions in space and assess sensitivity of variable selection approaches to model specification. The proposed models were fitted on malaria data from two national surveys in Angola to identify the best proxies of intervention coverage measures on malaria risk and find the provinces in the country that interventions have an important effect on the disease.
In Chapter 6, we develop a computational algorithm that we called iteratively integrated nested Laplace approximations (i-INLA) to perform variable selection of spatiotemporally varying coefficients of non-Gaussian data via a marginal likelihood approximation. We implemented the algorithm on the Angola malaria data to assess effects of interventions in space and time on the dynamics of malaria. We use the predictive process approximation to the spatial components of the models to speed inference. Effects of the predictive process approximation on variable selection are investigated.
This PhD thesis contributes to the fields of Bayesian spatial modelling and spatiotemporal epidemiology of tropical diseases with: (i) methodology for Bayesian variable selection of spatiotemporally varying coefficients allowing flexible inference, especially for computationally intensive geostatistical models of data collected over large number of locations; (ii) sensitivity analysis of Bayesian variable selection formulations of models with spatially varying coefficients; (iii) estimates of incidence rates for cutaneous and visceral leishmaniasis in Brazil depicting the current situation of leishmaniasis in the country; (iv) an open-access georeferenced database cataloguing all available survey data for soil-transmitted helminth infections in sub-Saharan Africa and Cambodia for disease control and research purposes; (v) up-to-date smooth risk maps, and estimates of the number of people infected and of the required treatments of soil-transmitted helminth infections in sub-Saharan Africa and Cambodia; (vi) an evaluation of the predictive ability of cluster-aggregated WASH and other SES-related proxies in disease mapping of poverty-related diseases; and (vii) geostatistical models of malaria risk for estimating effects of malaria intervention coverage measures across space and over time.