Bayesian spatial-temporal modelling to assess the impact of climate change on the burden of malaria and to support adaptation tools in Burkina Faso

Malaria is one of the oldest infectious diseases and a major global health challenge. It is a vector-borne disease caused by the Plasmodium parasite, transmitted to humans through the bite of infected female Anopheles mosquitoes. Sub-Saharan Africa (SSA) bears the brunt of this burden, accounting for over 90% of global malaria cases and deaths. The high level of malaria transmission in this region is primarily driven by favorable climatic conditions, inadequate health systems, and the low socio-economic status of affected populations. Consequently, malaria contributes to a substantial loss of life in endemic countries, particularly among children under five and pregnant women, as well as economic setbacks due to lost workdays. These challenges further impede socio-economic development, perpetuating a vicious cycle of poverty in the affected countries.
Malaria is endemic in Burkina Faso, putting the entire population at risk. Significant progress has been made in reducing the burden of malaria following the scale-up of interventions over the past two decades. However, the country still ranks among those with the highest burden, and malaria remains the leading cause of outpatient visits to health centers across all age groups in 2022. This persistent high burden suggests that factors such as climate change may be counteracting malaria control efforts by expanding areas suitable for transmission.
While many statistical and mechanistic models have been developed to evaluate the impact of climate change on malaria burden, the resulting evidence remains inconclusive. Most models fail to account for the interactions between climate change, malaria interventions, and socio-economic factors, which are likely to influence future disease burden. This limitation is partly due to the scarcity of high-quality, long-term malaria data that include non-climatic variables. However, the health and demographic surveillance systems (HDSS) sites and the national health data warehouse of Burkina Faso routinely collect data on malaria incidence, mortality, interventions, and household-level indicators. The availability of these data provides an opportunity to model the spatial-temporal interactions between climatic and non-climatic factors on the burden of malaria. Strengthening model-based surveillance to better understand the interactions between climatic and non-climatic factors, map disease distribution and forecast disease incidence will improve preparedness for the impacts of climate change.
The overarching goal of this thesis is to deepen our understanding of the effects of climate and control interventions on the burden of malaria and to develop methods to strengthen disease surveillance and climate adaptation strategies. Specifically, the objectives are to: i) assess the relative effect of climatic factors and malaria control interventions on the spatial-temporal change of parasitaemia risk; ii) estimate the impact of climate variability and interventions on malaria incidence and forecasting; iii) assess the role of interventions and climate on malaria mortality among children under five; and iv) evaluate the impact of sampling weights on geostatistical malaria risk modeling.
In Chapter 2, Bayesian logistic geostatistical models were fitted on Malaria Indicator Survey (MIS) data collected in Burkina Faso during 2014 and 2017/2018 to estimate the effects of malaria control interventions and climatic factors on the temporal changes of malaria parasite prevalence. Additionally, intervention effects were assessed at the regional level using a spatially varying coefficients model. Night temperature showed a statistically important negative association with the geographic distribution of parasitaemia prevalence in both surveys, while the effect of insecticide-treated nets (ITNs) use was negative and statistically important only in 2017/2018. Overall, the estimated number of infected children under the age of 5 years decreased from 704,202 in 2014 to 290,189 in 2017/2018. The use of ITNs was related to this decline at both the national and regional levels, while coverage with artemisinin-based combination therapy (ACT) was only important at the regional level.
In Chapter 3, we assessed the impact of climate variability and interventions on malaria incidence and forecasting in Burkina Faso. We estimated time delays in the effects of climatic factors on malaria incidence and investigated how fluctuations in these factors affect malaria transmission dynamics across different time scales and three climatic zones (hot/arid, cooler/wet and moderate temperature zones). Monthly confirmed malaria cases in children aged under five years from 2015 to 2021, as recorded in the District Health Information System 2 (DHIS2) were modeled using Bayesian generalized autoregressive and moving average (GARMA) negative binomial models. Wavelet analysis was conducted to investigate the temporal dynamics of the transmission. During the study period, malaria incidence averaged 9.92 cases per 1,000 persons per month. The highest incidence was observed in July and October in the cooler/wet zone and October in the hot/arid and moderate zones. Periodicities at 6-month and 12-month intervals were identified in malaria incidence and Land surface temperature (LST) and at 12 months for rainfall from 2015 to 2021 in all climatic zones. The highest predictive power was observed at lead times of 3 months in the cooler/wet zone, followed by 2 months in the hot/arid and moderate zones. The forecasting ability varied across the zones. ITNs were not statistically important in the hot/arid zone, while ACTs were not in the cooler/wet and moderate zones.
In Chapter 4, Bayesian negative binomial temporal models were employed to assess the role of control interventions and climatic factors on malaria mortality among children under five, using two decades of data from the Nouna HDSS (2002-2021). We further evaluated the role of climatic seasonality on the seasonality of malaria deaths. The results showed that the lag time in the effects of climatic factors varied over time. Malaria mortality, rainfall, and LST exhibited a 12-month seasonal cycle throughout the years, while LST also showed a 6-month cycle in specific years. Rainfall lagged by 1.5 to 2 months and LST by 1 to 1.5 months, depending on the seasonal cycle and year. Rainfall was associated with a 59% increase in malaria mortality, while LST was linked to a 32% decrease. ITN use was associated with a 41% reduction in mortality, but ACT coverage was not statistically important.
In Chapter 5, a Bayesian logistic geostatistical model was fitted to the Burkina Faso 2021 demographic and health survey (DHS) data to assess the impact of sampling weights on inferences from geostatistical malaria risk modeling. We estimated the effects of climate factors and malaria interventions on the geographical distribution of prevalence. A spatially varying coefficient model was also used to evaluate the effects of interventions at the provincial level. We compared inferences from these models to those from their counterparts that adjusted for sampling weights. The results revealed that the model adjusted for sampling weights provided a better fit, with a lower deviance information criterion (DIC) of 1185.41 compared to 1365.18 in the unadjusted model. Spatial variance was reduced in the adjusted model (0.30, 95% Bayesian credible interval (BCI): 0.16-0.48) compared to the unadjusted model (0.83, 95% BCIs: 0.39-1.49), and the spatial range was shorter at 44.4 km versus 108.46 km. Both models identified negative associations between malaria prevalence and factors such as ITN coverage, LST, night-lights, and distance to permanent water, with stronger effects and shorter BCIs in the adjusted model. The unadjusted model underestimated predicted prevalence, the number of infected children, and population-adjusted prevalence, while overestimating intervention effects across provinces.
The results from this thesis contribute to a better understanding of the impact of climate and control interventions, as well as the interactions between these factors, on malaria burden. The findings could inform the development of more effective malaria control strategies in Burkina Faso by enabling policymakers to gain deeper insights into malaria burden within specific climatic zones and allocate interventions accordingly. The estimated risk and intervention effect maps serve as invaluable resources for identifying high-risk areas and regions where intervention efficacy is limited. Moreover, the forecasting models can be adopted by national malaria control programs (NMCPs) for both short- and long-term predictions. Additionally, our findings contribute to the improvement of statistical methods for estimating malaria parasitaemia risk when using complex health survey data, such as that collected through the MIS or DHS.