Modelling the seasonal and spatial variation of malaria transmission in relation to mortality in Africa

About three billion people worldwide are estimated to be at risk of malaria transmission. In developing countries, malaria is believed to be a major cause of morbidity and mortality, mostly in children under five years. It is among the indirect causes of maternal mortality and infants’ deaths due to low-birth-weights. Malaria brings huge economic burden due to number of days lost during sickness and deaths, sustaining a vicious cycle of disease and poverty in sub Saharan Africa (SSA) and high attribute of disability-adjusted life years.
A number of malaria control interventions to reduce intensity of transmission have been successfully implemented in the regions of SSA, however, elimination of malaria is still a dream in many developing countries today. Failures in global eradication are related to resistance in insecticides and anti-malarial drugs, and health systems related factors. The Roll Back Malaria (RBM) partnership reinforced new strategies to combat malaria with long-term goal of eradicating the disease globally. This was facilitated by increasing funding for malaria research, improve multi disciplinary initiatives and make malaria among the main agenda of all international health and development forums. The reduction in mortality, especially in children has been reported recently and is associated with achievements in intervention strategies, improvements in malaria diagnosis and treatment. However, poor natural acquisition of malaria immunity in children as a consequence of weak or no exposure is a major epidemiological concern and brings a fear of higher mortality rates or shifting of age of death to older children. Understanding and quantify links between transmission, intervention, immunity and mortality is key for sustainable progress towards malaria control targets.
A comprehensive analysis of information on malaria transmission, vital events, drivers of transmission and mortality-related risk factors is required to achieve that. Lack of vital registration systems in developing countries hinders availability of appropriate data to conduct such analysis. Establishment of Demographic Surveillance Systems (DSS) in many developing countries aims to fill these information gaps. One of the initiatives integrated within DSSs is the Malaria Transmission Intensity and Mortality Burden across Africa (MTIMBA) project. The project compiled a database of mosquito collections at selected sites in Africa over a large number of locations, using standardized methodologies for a period of three years. The entomological parameters were linked with routinely monitored vital events within the DSS. The MTIMBA database is the most comprehensive entomological database ever collected in Africa which allows studying spatial-temporal variation in malaria transmission in relation to mortality.
Malaria is an environmental disease hence transmission varies with climate as it modifies population, survival, distribution and infectivity of malaria vectors. Quantification of association between climate and transmission is important to allow prediction of risk even in areas that field data cannot be easily obtained. Development in geographical information systems (GIS) and availability of remote sensing (RS) data facilitates availability of environment and climate data at high space and time resolutions allowing accurate estimation of outcome-factor relationship.
However, DSS data are large, sparse, zero-inflated and are characterized by seasonal patterns, spatial and temporal correlations. Standard models assume independence between observations, an assumption which do not hold for correlated data, hence utilizing these models might result into biased estimates. Geostatistical modeling of large, sparse and zero inflated space-time data is computational challenging specifically in the estimation of the spatial processes. The spatial correlation is accounted by introducing location-specific random effect parameters which are assumed to arise from a spatial process quantified by a multivariate normal distribution. The models are highly parameterized and their fit is computationally intensive. Bayesian computational algorithms such as Markov Chain Monte Carlo (MCMC) can be used to fit these models. Estimation of the spatial process requires inversion of the covariance matrix at each simulation point. The dimension of the matrix increases exponentially with number of locations and the inversion becomes infeasible when the size is too large. Recent techniques overcome this problem by approximating the spatial process from a subset of locations. These methods have been applied on Gaussian outcomes observed over a grid. Extension and formulation of rigorous methods to efficient model MTIMBA data are needed to allow precise prediction of malaria transmission at locations with mortality data to enhance studying the association. Lastly, seasonality in climatic conditions which introduces seasonal patterns in transmission and mortality data, should be accounted for when modelling such data.
The objectives of this thesis were to i) develop Bayesian geostatistical models to analyze very large and sparse geostatistical and temporal non-Gaussian data with seasonal patterns and ii) apply these models to (a) estimate space-time heterogeneity in malaria transmission (b) assess mortality variations between different ages during the first year of life while adjusting for seasonality and (c) determine the relation between transmission intensity and risk of mortality in children and adult population after taking into account control interventions. This work used an extract of MTIMBA data from the Rufiji DSS (RDSS) collected between October 2001 and September 2004.
Evaluation of approaches to capture seasonal pattern is discussed in Chapter 2 and applied to estimate mortality peaks at different stages of infant life. In Chapter 3, models approximating the spatial process from a subset of locations were developed to assess effect of climate, seasonal and spatial pattern of sporozoite rate (SR) of An. funestus and An. gambiae in RDSS. A rigorous approach to analyze malaria transmission data using Entomology Inoculation Rate (EIR) data, which is the product of mosquito density and SR, is discussed in Chapter 4. Zero-inflated models were used to account for over-dispersion and zero-inflation in the data. High resolution EIR estimates were produced for the RDSS. Exposure surfaces obtained in Chapter 4, were aligned with mortality events to assess the relationship between all-cause mortality and malaria transmission. Geostatistical Bernoulli discrete-time regression models adjusted for age and ITN possession were used for that analysis. The results of these analyses are presented in Chapters 5 and 6. The EIR was incorporated in the model as a covariate with measure of uncertainty.
This work is a building block on the insight and understanding of association between malaria transmission and all-cause mortality. The strength of results of this work relies on EIR estimates predicted at high spatial (household level) and temporal resolution by employing rigorous geostatistical models fitted on large entomological data. The better exposure estimates obtained are able to more accurately estimate the mortality-transmission relation.