Geostatistical modelling and survey sampling designs for malaria control and surveillance

In many low-and middle-income countries, malaria is endemic and remains a serious health threat and a significant contributor to mortality. According to the World malaria report 2019, more than ninety percent of malaria cases occurred in Africa. Children remain the most vulnerable group. Cameroon belonged to the list of countries most affected by malaria. The country is subdivided into several ecological zones in which malaria prevalence is spatially heterogeneous. The disease transmission is highly seasonal in the North, perennial in the southern and eastern parts and relatively low in the high mountains of West and Adamawa regions.
Substantial efforts made by international donors and the Cameroon government led to a decline in the malaria burden over the last decade. Many health interventions and actions were implemented to fight against the spread of disease. National surveys such as the Demographic and health surveys (DHS), malaria indicator surveys (MIS), and multiple indicator cluster surveys (MICS) are conducted every three to five years collecting individual and household level data on malaria, disease interventions and socio-economic factors to measure the progress achieved in the control of the disease. The data are georeferenced at the centroid of clusters consisting of groups of around 25 households. For confidentiality reasons the cluster coordinates are reported with jittering, that is they are misplaced within a buffer around the actual location. Bayesian geostatistical models are commonly used to predict the geographical distribution of disease prevalence and quantify the effects of interventions. Predictions are improved by including in the models climatic and environmental proxies available at high spatio-temporal resolution from remote sensing sources. Despite the large number of survey data that have been analysed using geostatistical models, there are few studies of the effects of survey design factors on model-based estimates.
The overall goal of the thesis is to evaluate and further strengthen the methodology of the malaria survey designs used for disease monitoring and evaluation, in particular the aspects related to the timing of the survey, the jittering in the coordinates of the reported cluster locations, and the seasonal monitoring of malaria data. Furthermore, we evaluated the ability of the survey data to estimate the malaria-related deaths and compared estimates of the effects of malaria interventions using data from malaria surveys and the Health Management Information System. More specifically, the thesis pursues the following objectives: i) assess the influence of the survey timing by comparing the DHS and MIS geostatistical model-based malaria risk estimates obtained at different malaria transmission seasons; ii) assess the effects of the DHS jittering algorithm on the prediction of malaria risk and on the estimates of the disease risk factors; iii) evaluate the effects of malaria interventions on the geographical distribution of disease incidence after adjusting for the effects of climatic and environmental factors; iv) improve malaria disease and vector survey sampling designs by optimizing the selection of survey locations and v) evaluate the ability of malaria surveys to estimate malaria-related deaths.
The above objectives were addressed by analysing DHS, MIS and MICS survey data from Cameroon as well as malaria incidence data from the Health Management Information System. The methodology and results for each objective are included in five main chapters.
In Chapter 2, Bayesian stationary geostatistical model are employed to analyse MIS and DHS, national surveys conducted in 2011 during the rainy and dry seasons, respectively. Geostatistical variable selection was applied to identify the most important climatic factors and malaria intervention indicators. The results showed that the timing of the malaria survey influences estimates of the geographical distribution of disease risk, especially in settings with seasonal transmission. In countries with different ecological zones and thus different seasonal patterns, a single survey may not be able to identify all high-risk areas.
Chapter 3 presents the influence of jittering on the assessment of intervention effects and the spatial estimates of disease risk distribution at high spatial resolution. Based on original MIS cluster locations, a set of a hundred (100) shifted cluster locations were generated using the DHS jittering algorithm. Bayesian variable selection applied in the original dataset locations as well as for each one of the hundred jittered datasets to select climatic/environmental predictors, socio-economic factors and malaria intervention indicators. Geostatistical models were applied to the original as well as the simulated data using the selected covariates. The results indicated that the selections of important climatic predictors and of intervention indicators were influenced by the jittering, while estimates of the disease risk at high geographical resolution were slightly affected.
Chapter 4 focused on the selection of relevant cluster locations for an efficient assessment of intervention effects. Based on MIS data, a Bayesian geostatistical model was applied to the most important climatic predictors to estimate the malaria risk, and associated uncertainty over a high resolution gridded surface across Cameroon. An adaptive algorithm was proposed to select survey locations based on a multi-criteria objective approach. The adapted algorithm was able to identify a meaningful subset of cluster locations based on their contribution to the uncertainty and the needs of the national malaria program.
In chapter 5, we evaluated the effects of interventions on the spatiotemporal dynamic of malaria incidence and the capability of HMIS data to capture disease pattern. From 2012 to 2016, confirmed malaria data were extracted and aggregated by month at the district level. During the same period, climatic factors were obtained from satellites and averaged over the district surface. Bayesian variable selection was applied to identify the most important lag time for each continuous climatic factor. A Bayesian spatiotemporal of the relationship between malaria incidence and intervention was fitted and adjusted with the important climatic factor. The percentage of households having one ITN per two persons was identified as the most important coverage indicators, while the normalized difference vegetation index, rainfall estimates were selected among climatic predictors. Having an ITN for every two persons was negatively associated to malaria cases. The incidence maps drawn at district level were able to capture patterns of disease risk that were not estimated by the DHS and MIS data.
Chapter 6 assessed the relationship between malaria prevalence and all-cause mortality in infants and in children under-5 years old, by considering seasonal influence of malaria transmission as well as socio-economic factors. Bayesian geostatistical Bernoulli and zero-inflated Bernoulli models were fitted on the mortality risk data. A statistically important relation was estimated between infants (excluding neonates), under-five years old mortality and malaria risk. The effects of malaria parasite risks on under-five mortality became more statistically important in the absence of neonates. Mortality in the under-five group was reduced during the dry season.