Study area
The study was conducted in the Greater Accra Region, which includes the country’s capital city, Accra (Figure 1). It is the smallest of the 16 administrative regions in Ghana. The area is about 3,245 square kilometers and it is the second-most populated region in the country with a population of 5,055,765 in 2019 [36]. It is divided into 29 local government areas (LGAs). Most of the LGAs are urban and urban LGAs house more than 80% of the region’s residents [37]. The LGAs are classified as districts, municipalities and metropolitan areas, with metropolitan areas being the most urban. However, for the purpose of this study, they are all referred to as districts. The blend of remote and urban communities within the region makes it a suitable site for this study. The climate of the area is categorized as tropical savannah.
Data source
The study used secondary aggregated clinical data for all age groups, including both in-patients and outpatients with laboratory-confirmed malaria infections from 2015–2019 in the Greater Accra Region. A total of 1,105,370 malaria cases were identified through passive case detection at health facilities and collated in the DHIMS at the district level. Yearly population data for each district were obtained from the Health Information Department of Greater Accra Regional Health Directorate.
Long-term average annual and seasonal temperature and rainfall variables were determined using data obtained from the WorldClim project at a spatial resolution of 1 km [38]. The variables obtained from WorldClim had been created by spatial interpolation of climate data gathered from global weather station sources between 2010 and 2018 utilising a thin=plate smoothing spline algorithm.
Polygon shapefiles of administrative boundaries at the district level of the Greater Accra Region were obtained from the DIVA-GIS website ( www.diva-gis.org). The spatial datasets including Standard Morbidity Ratios of malaria cases were imported into ArcGIS version 10.7.1 (ESRI Inc. Redlands, CA, USA, URL: https://www.esri.com) and projected to the Universal Transverse Mercator (UTM) coordinate system (zone 48 N).
Analysis of seasonal and inter-annual patterns
The mean monthly number of malaria cases was calculated from the full-time series (January 2015–December 2019). Seasonal-trend decomposition, based on locally (STL) weighted regression was used to decompose the time series of malaria incidence to reveal the seasonal relationship, inter-annual pattern, and the residual variability. The STL model was structured as follows:
where Yt, St, Tt and Rt represent the local malaria cases with logarithmic transformation, additive seasonal component, trend, and remainder component respectively while t signifies time in months [35, 39, 40].
Standardized morbidity ratios
Standardized morbidity ratios (SMRs) per district were analysed using the following formula:
Where Y denotes the total SMR in district i, O and E are the total number of the observed and expected malaria cases in district i across the study period. The expected number (E) was calculated by multiplying the regional malaria incidence by the average population for each district over the study period.
Annual parasitic incidence
Annual parasitic incidence (API) per district were calculated using the formula:
Independent climatic variable selection
A preliminary Poisson regression was used to select the significant climatic covariates. Maximum and minimum temperature, and rainfall with zero, one, two, three, four, five and six-month lag times, were entered into univariate Poisson models. Significant (p<0.05) climatic variables with lag times with the lowest Akaike’s information criterion (AIC) were selected for inclusion in the model (Supple Table 1). The co-linearity of selected variables was tested using variance inflation factors (VIF) (Supple Table 2). Minimum temperature without lag, rainfall lagged at one month and maximum temperature lagged at six months were selected. Preliminary statistical analyses were all performed using the STATA software version 16.0 (Stata Corporation, College Station, TX, USA, URL: https://www.stata.com).
Spatio-temporal model
Poisson regression models for malaria cases were created using the Bayesian statistical software WinBUGS version 1.4 (Medical Research Council, Cambridge, UK and Imperial College London, UK). Three models were created incorporating, spatially unstructured (Model I), spatially structured (Model II) and both structured and unstructured random effects (Model III). Each model included the climatic variables as fixed effects. The best-fit parsimonious model was selected with the lowest DIC. Model III, which includes all components of the other models was structured as follows:
where Y are the observed counts of malaria, for ith district (i=1…60) in the jth month (January 2015 to December 2019), E are the expected number of malaria cases included as an offset to control for population size and θ is the mean log relative risk (RR); α is the intercept, and β1, β2, β3, and β4 are the coefficients for monthly malaria trend, rainfall lagged at one month, maximum temperature lagged at six months and minimum temperature without lag. The unstructured and spatially structured random effects were represented by ui and si, each with a mean of zero and with variances of σu2 and σs2 and wij is the spatiotemporal random effect (with a mean of zero and variance of σw2).
The spatially structured random effect was calculated using a conditional autoregressive (CAR) prior structure. Spatial relationships between the districts were computed using queen contiguity, where an adjacency weight of 1 was allocated if two districts have a common border or vertex and 0 if they did not. The intercept was delineated with a flat prior distribution, while the coefficients were defined by a normal prior distribution. Non-informative gamma distributions characterise by shape and scale parameters equivalent to 0.01 were used to specify priors for the precision (1/ σu2 and 1/ σs2) of the unstructured and spatially structured random effects. Additionally, models were established without the structured (Model I) and unstructured (Model II) random effects to determine if including them improved model fit.
The burn-in, comprising the initial 10,000 iterations, were discarded. The simulation chains were then run for blocks of 20,000 iterations to assess for convergence. Convergence was determined through visual inspection of posterior density and history plots for each model and was achieved at 100,000 iterations. Markov Chain Monte Carlo simulation with Gibbs sampling was used to estimate model parameters [41]. Values of the posterior distributions were then stored and summarised for analysis (posterior mean and 95% CrI).
An α-level of 0.05 was used to indicate statistical significance (as shown by 95% CrI for coefficients (β) that excluded 0). ArcMap 10.7.1 software (ESRI, Redlands, CA, URL: https://www.esri.com) was used to produce maps of the posterior means of the random effects from the three models.
Ethical considerations
Ethical approval for the study was obtained from the Human Research Ethics Committee of the Australian National University (Protocol: 2020/465). Permission for the study was also granted by the Ethics Review committee of the Ghana Health Service (protocol: GHS-ERC 051/07/20). The study was conducted in accordance with guidelines and regulations from the above two ethic committees regarding the use of secondary data. The Greater Accra Regional Health Directorate also approved and provided the malaria dataset for the conduct of the study within the region. The dataset provided did not include any personally identifying information and could not be linked back to study participants by the authors.