Developing Bayesian spatio-temporal models of assess the relation between malaria transmission and mortality in Burkina Faso

Malaria control remains a major public health challenge especially in sub-Saharan African countries. In spite of the rapid decline observed in malaria mortality in Africa over the last decade due to scaling up of control interventions and social/economic development, malaria mortality figures remain unacceptably high.
An estimated 198 million cases of malaria worldwide led to nearly 584,000 deaths in 2013. The majority of the illnesses (85%) and the case fatalities (90%) occur in Africa taking its greatest toll among young children under five years of age. Beside the deaths toll, repeated clinical malaria episodes cast an enormous economic burden on households.
Predicting the effectiveness of malaria interventions at a given place requires appropriate information on both mortality and transmission levels in order to ascertain the level of efforts required to achieve a significant reduction in morbidity as well as the number of deaths that could be prevented.
This quantitation is needed for estimating the burden of the disease based on different transmission levels and for building models, which incorporate this relationship in order to predict the likely effects of malaria interventions on mortality. Yet, for many of the sub-Saharan countries, most severely burdened by malaria, these crucial estimates are lacking making it difficult to accurately predict the likely impact of malaria interventions on mortality.
The Malaria Transmission Intensity and Mortality Burden Across Africa, INDEPTH-MTIMBA project was initiated in the 2002 in a number of Health and Demographic Surveillance Systems (HDSS) sites. HDSS are sites that are routinely monitor all life events in a certain area and are used for estimating mortality in the absence of complete registration of deaths and births in many developing countries. The MTIMBA project aimed at assessing the levels of malaria transmission intensity, establishing the relationship between all-cause, malaria-specific mortality and malaria transmission intensity taking into account the effect of disease control interventions. MTIMBA collected georeferenced entomological data, biweekly during a period of 3-4 years. One of the HDSS sites of the MTIMBA project was Nouna in Burkina Faso, however data have not yet been analysed. Previous studies have tried to assess the relation between malaria endemicity and mortality using mortality data from the Demographic Health Surveys (DHS) and malaria data from Malaria Indicator Survey (MIS). In Burkina Faso, the DHS-Multiples Indicator Cluster Survey (BFDHS-MICS) of 2010 was the first survey that collected georeferenced data of both child mortality and malaria endemicity across the country.
The overall goal of the thesis is to assess the association between malaria transmission and mortality at different geographical scales in Burkina Faso. The specific objectives of the research are to (i) obtain time- dependent and spatially explicit estimates of entomological inoculation rate (EIR) within the Nouna HDSS site; (ii) obtain spatially explicit estimates of malaria parasite risk, number of infected children and assess the effects of malaria interventions in Burkina Faso; (iii) assess the relation between infant and under-five mortality and malaria endemicity in Burkina; (iv) assess the relationship between malaria transmission and mortality (all-cause and malaria-specific) across different age groups in Nouna HDSS and (v) assess the ability of verbal autopsies to diagnose malaria as a cause of death using the malaria-transmission relation as a gold standard. We addressed the above objectives by employing Bayesian spatio-temporal models and analysing the Burkina Faso DHS (BFDHS-MICS) 2010, the MTIMBA and the mortality databases from the Nouna HDSS site.
In chapter 2, the MTIMBA data were analysed to obtain surfaces of malaria transmission across the Nouna HDSS. In particular, Bayesian geostatistical zero-inflated binomial and negative binomial models including harmonic seasonal terms, temporal trends and climatic remotely sensed proxies were applied to assess spatio-temporal variation of sporozoite rate and mosquito density in the study area. Bayesian variable selection was applied to determine the most important climatic predictors and elapsing (lag) time between climatic suitability and malaria transmission. Bayesian kriging was used to predict mosquito density and sporozoite rate at unsampled locations. These estimates were converted to covariate and season-adjusted maps of entomological inoculation rates. The results showed that Anopheles gambiae is the most predominant vector (79.3%) and is more rain-dependant than its sibling Anopheles funestus (20.7%). Variable selection suggested that the two vector species react differently to different climatic conditions. Prediction maps of EIR depicted a strong spatial and temporal heterogeneity in malaria transmission risk despite the relatively small geographical extend of the study area.
In chapter 3, Bayesian geostatistical models and BFDHS-MICS 2010 survey data were used to assess the effects of health interventions related to insecticide-treated nets (ITNs), indoor residual spray (IRS), artemisinin-based combinations therapy (ACT) coverage associated with childhood malaria parasite risk at national and sub-national level after taking into account geographical disparities of climatic/environmental and socioeconomic factors. Several ITN coverage measures were calculated and Bayesian variable
selection was used to identify the most important ones. Parasitaemia risk surfaces depicting spatial patterns of infections were estimated. The results showed that the population adjusted predicted parasitaemia risk ranges from 4.0 % in Kadiogo province to 82% in Kompienga province. The effect of ITN coverage was not important at national level; however, ITNs had an important protective effect in Ouagadougou as well as in three districts in the western part of the country with high parasitaemia prevalence and low-to- moderate coverage. There was a large variation in ACT coverage between the districts. Although at national level the ACT effects on parasitaemia risk was not important, at sub-national level, 18 districts around Ouagadougou delivered effective treatment.
In chapter 4, we used data form the Burkina Faso first nationally representative household survey focusing on malaria-related indicators, BFDHS-MICS 2010 and apply Bayesian geostatistical Weibull survival models to explore the relationship between malaria and infant/child mortality in Burkina Faso after adjusting for, both individual child and household or family characteristics as well as mother’s birth history. There is a significant relationship between malaria endemicty and child survival in urban settings. Children living in the urban settings with endemicity level above 75% are at higher mortality hazards. Other predictors of infants and child survival are those related to biological (birth size, mother age at birth), demographic socioeconomic and antenatal care factors.
In chapter 5, we used entomological data, which, were collected biweekly from 2001-2004, and mortality data extracted from the Nouna HDSS database. We address spatial misalignment between the two data sources by obtaining EIR estimates at the mortality locations using Bayesian spatio-temporal models. Analyses were adjusted for socioeconomic status (SES) and ITN coverage. Time to death was treated at monthly interval and Bayesian geostatistical logistic regression approximating Cox proportional hazard model and incorporating the predicted EIR as covariate with measurement error were fitted. The mortality rates were highest in 2001, 17 (95% CI: 15.1, 19.1) and 2003, 13.8 (95% CI: 12.95, 14.8). The overall mortality rate over the study period was 11.3 (95% CI: 10.8, 11.7). The highest mortality rates were observed in children and old age groups with the respective rates of 23.9 (95% CI: 22.4, 25.4) and 81.9 (95% CI: 75.8, 88.5). A positive log-natural relationship between mortality and EIR was found among children (1-4 years), while a protective effect was found among adolescents/adults (15-59 years). The highest mortality risk associated with EIR was observed among children (5%).
In chapter 6, we used the same approach as in the previous chapter however focusing the interest on malaria specific mortality in order to assess the relationship between malaria specific mortality and EIR within the Nouna MTIMBA-HDSS site. The sensitivity and specificity of the physician-certified verbal autopsy (PCVA) were also assessed. Results showed that malaria mortality rates were highest in years 2001, 5.4 (95% CI: 4.4, 6.6) and 2003, 4.1 (95% CI: 3.6, 4.7). A significant positive natural logarithmic relationship was found between malaria exposure and mortality among children, with hazard ratio (HR) of 1.06(95%CI:1.03,1.08). Thepercentageofdeathsassigned-malariaascauseinVAwashighestinchildren and adults with respectively 45% and 35.3%. The percentage of deaths attributable to malaria exposure was in old-age group (93.9%). The overall specificity of the PCVA is 0.70.
Results of this work contribute to a better understanding of the interplay between environmental/climatic conditions and malaria transmission, which is important not only for delivering interventions at the right time but also for developing predictive models to support early warning systems (EWS).
The estimated risk and intervention effect maps are valuable tools for identifying high-risk areas and areas with less effective interventions in order to improve malaria control in Burkina Faso. These outputs can serve as benchmarks to evaluate the effectiveness of future control interventions and progress of the efforts towards disease control.
Results from the mortality-malaria transmission analyses improve our understanding of the relationship between malaria transmission, all-cause and malaria specific mortality in Nouna region.