Geo-visual integration of health outcomes and risk factors using excess risk and conditioned choropleth maps: a case study of malaria incidence and sociodemographic determinants in Ghana

Ghana, located in the Sub-Saharan African (Figure 1), has a population of more than 27 million people and a population density of 113persons/km². Approximately 51% of the populations are residing in urban centres according to the 2010 census [16]. Ghana has ten administrative regions, subdivided into 170 districts on which 2010 census was conducted, and has approximately 75 ethnic groups with different socio-cultural practices [16, 17]. This study was conducted on the 170 districts using malaria incidence for the 2010 – 2014 period.

Clinically diagnosed malaria cases for outpatient visits at all health facilities in Ghana during the study period 2010–2014, were obtained from the Centre for Health Information and Management (CHIM) within the Ghana Health Service (GHS). Routinely, clinical diagnoses of malaria are based on parasitological microscopy and/or rapid diagnostic test (undertaken at public and private hospitals, clinics and health centres) but mostly by rapid diagnostic test at Community Health Planning Services programme (CHPS) zones in accordance with World Health Organization criteria. The CHPS programme improves coverage of malaria ascertainment for underserved communities and villages in rural areas by using trained community health nurses to render basic clinical and public health services, including diagnosis and treatment of malaria. The entire population of Ghana is at risk of malaria since malaria is endemic in all parts of the country with seasonal variations. Shapefiles for the 170 local health administrative districts were obtained from the Survey and Mapping Divisions, Accra and the Geomatic department of KNUST. Data were available nationally at the district level. Sociodemographic characteristics were obtained from the 2010 Population and Housing Census (PHC) which had complete population coverage on the 170 districts, providing information relating to the various aspects of the populations and households. The district-level proportions (expressed as percentage of the total) of the socio-demographic factors used in the study were described concisely as follow:

Malaria incidence rates were estimated followed by assessment of the spatial dependency within and between the health outcome (malaria) and risk factors (sociodemographic risk factors). Statistically significant risk factors were selected for the excess risk and conditioned choropleth maps.

The crude district-level annual malaria incidence rates, R_Mal for the i-th district in the year t was estimated as

where\( {X}_{Mal_{it}} \) denotes the reported malaria case counts at the district i (i= 1, 2, … , n =170) for the year t (t = 2010, 2011, … , 2014), andP_it denotes the population in district i for the year t. The cumulative and five-year average incidence rates were also calculated for each district.

A spatial weights matrix was created based on first-order queen polygon contiguity. The effects of first-order queen polygon contiguity, merging both rook and bishop contiguities, are sufficient to capture spatial autocorrelation given the size and shape of the districts in Ghana. The irregularity of the shapes of the districts, hence the adoption of this contiguity approach in past studies to avoid neighbourless districts [19‐21] and it is suitable to represent malaria transmission. Rook or bishop contiguity can leave gaps, which would not represent malaria transmission very well. Hence districts that shared common edges and/or common corners were considered neighbours and weights were assigned to these identified neighbours. The spatial weights were row-standardized such that for each row Σw_ij = 1 if districts i and j shared a common boundary; otherwise Σwij = 0, for non-neighbouring districts. Following standard convention, we excluded “self influence” by assuming that w_ii = w_jj = 0 so that W has zero diagonals.

We used Empirical Bayes Smoothing using the principle of shrinkage [1, 22, 23] to stabilise incidence rates for areas with small populations or disease counts. We assumed that the relative risks of people residing in district i

(δ_i) were independently and identically distributed according to a Poisson distribution:

where x_i is the random variable representing disease count in district i while N_i is expected count for the same district. The Empirical Bayes Smoothed (EBS) relative risk of malaria, \( {\widehat{R}}_{Mal_{it}} \) borrows the neighbouring district rates to adjust the uncertain rates as per the expression:

where ϕ_i is the ratio of prior variance to the data variance, and \( {m}_{\delta_i} \)is the prior mean (weighted sample mean). The final EBS rate remains practically unchanged for districts with relatively large population or cases [23].

We checked and established spatiotemporal patterns of rates from 2010 to 2014 using global and local Moran’s indices. Global Moran’s I was used to determine whether or not identifiable spatial patterns exist over space and time [4, 5, 22] and Anselin local Moran’s I_i (the most widely used Local Indicator of Spatial Association, LISA), to identified specific districts and locations exhibiting spatial autocorrelation with their neighbouring districts as clusters or outliers [23‐25]. The statistical inference was based on Monte Carlo randomisation test at 999 permutations with significance pseudo p-value<0.05 [4, 5, 19]. Non-spatial correlation was evaluated with Pearson correlation while global bivariate Moran's I was estimated to examine the spatial correlation between the five-year average incidence of malaria and the sociodemographic covariates. The statistically significant sociodemographic determinants were selected for excess risk and conditioned choropleth maps. Due to the ERM and CCM computational functionality in GeoDa, all spatial statistical maps were generated using GeoDa software version 1.12 even-though this package has lower cartographic quality as compared to other spatial packages especially ArcGIS.

The excess or relative risk is a form of standard morbidity or mortality rate (SMR) often used in public health which is estimated as the ratio of observed rate to the expected rate. The expected rate is the average rate for all the population at risk in each location which is computed as the ratio of the sum of all events in all locations to the sum of all the populations at risk [4, 5]. Implemented with excess risk map functionality in GeoDa, we calculated excess risk maps (ERMs) of malaria incidence (event variable) for each statistically significant socio-demographic covariate (base variable) [4, 5, 26].

Both non-spatial Pearson correlation and spatial bivariate Moran's I analyses were performed between every pair of statistically significant risk factors to determine how they might act together or in sequence to influence malaria transmission. These analyses informed selection of the pairs of risk factors for the conditional choropleth mapping. We adopted conditioned choropleth mapping using the five-year average incidence rates of malaria as dependent variable (theme variable) and two strongly correlated significant sociodemographic factors (covariates) to visualise the three variables simultaneously. This resulted in a 3 x 3 panel of nine micromaps for which panel columns corresponded to the three categories of one covariate and the rows correspond to the three categories of the other covariate.

The minimum five-year average of 157 per 10,000 populations was observed for the capital city (Accra metropolis) of Ghana (Table 1). Sekyere East district recorded the greatest average incidence of malaria over the five-year period. The annual mean incidence of malaria in Ghana almost doubled during the study period from 996 per 10,000 in 2010 to 1,843 per 10,000 in 2014.

Natural breaks (Jenk’s) classification technique was used to identify categories of malaria incidence during the study period (Figure 2). Incidence of malaria varied geographically across the country with the greatest endemic districts located in the uppermost regions of the country and spread sparsely over the middle and southern belts.

Malaria incidence by district and year positively correlated with incidence in the neighbouring districts and the incidnece of the immediate previous year (Figure 5 in Appendix). This space-time association was strongest from 2012 to 2013 which exhibited the greatest Moran’s I and lowest pseudo p-value (<0.001). The local spatial maps indicated specific districts experiencing the high rates as compared to their neighbouring districts and further classify areas as clusters of high-high (called hotspot) and low-low (called coldspot) or outliers of low-high and high-low rates (Figure 6 in Appendix). On avearge, 143 out of the 170 districts had non-significant spatial clusterings while 7 and 14 locations were respectively identified as hotspot and coldspot clusters with the remaining 6 areas as outliers. The few districts with significant hotspots were located in the uppermost eastern part of the country. The results indicated that majority of the districts had malaria incidence rates not dissimilar from their neighbouring districts.

Among the 13 risk factors analysed, 7 of them had significant spatial autocorrelation (Table 2 in Appendix). The risk factor most spatially autocorrelated was the proportion of people in traditional African religious practices (Moran's I = 0.0657, z = 10.855). Seven socio-demographic determinants correlated significantly with the incidence of malaria either spatially and/or non-spatially. Urbanisation exhibited only non-spatial correlation and intramigration exhibited only spatial correlation, while basic education, none/other religion, intermigration, employment-to-population and proportion of household into farming correlated with malaria incidence both spatially and non-spatially. With the exception of intermigration and proportion of household in agriculture, all the significant determinants correlated negatively with the incidence rate of Malaria infection. Specifically, increases in the proportion of the population attaining basic education associated with decreases in malaria incidence. Districts that were more urbanised had lower incidence of malaria. Districts with greater proportions of people aged 15 years and over who were employed had lower incidence of malaria. Compared to no religious affiliation, christian, traditional and islamic religions had lower malaria incidence. The strength of the correlations for all covariates were weak with spatial correlations weaker than the non-spatial correlations.

Excess malaria morbidity maps were created using the statistically significant sociodemographic factors as the base covariate and the five-year average incidence of malaria as event variable. We did not derive an excess risk map for malaria with respect to urbanisation because some districts (N=7) were completely rural (urban population <5,000). Consequently, all the six risk factors used for the malaria excess risk analysis (Figure 3) had spatial correlation with the incidence of malaria. A greater proportion (116; 68.2%) of locations had more than expected malaria incidence using intermigration as the base covariate and where as high as 77(45.3%) of the districts had elevated rates that were more than four times as the expected rate. The majority of the districts (115; 67.6%), mostly at the south-western parts of the country, experienced more than expected incidence of malaria as compared to the average rate across the entire study area when intramigration was imposed as the base covariate. The distribution of excess malaria incidence for agriculture located 106 (62.4%) districts with incidence greater than expected. For influence of the employment to population ratio, most of the locations (61.8%) experienced more than expected incidence of malaria as compared to the average rate in the whole study area. For basic education, a little above half, 91(53.5%) of the districts had greater than expected malaria incidence, almost all of them clustered in the middle part of the country. Upon considering non-affiliation to religious groups as determining factor of malaria morbidity across the study, slightly above half of the districts (89; 52.4%) had malaria incidence greater than expected, and clustering was pronounced at the northern belt.

The strongest positive non-spatial correlation (p< 0.01, r = 0.969) was observed between Urbanisation and Basic education attainment (Table 3 in Appendix). The strength of spatial correlations (bivariate Moran’s I) were generally lower than the non-spatial Pearson correlations.

The significantly correlated paired covariates were used as conditioning variables (horizontal and vertical variables) of the average malaria morbidity (response or theme variable) to construct the malaria conditional choropleth micromaps (CCMs). The districts with revealing multivariate spatial relationship were indicated in the brown colour code and the degree of malaria incidence was expressed by the intensity of the colour and categorized into five, indicating very low, low, moderate, high and very high rates. Considering urbanisation-basic education- malaria CCM (Figure 4), when both urbanisation and basic education attainment were high (top-right panel), the emerging districts with high rates of malaria were found for the southern districts but not as many as when urbanisation and basic education were low (bottom-left panel), which were detected at the northern districts. The lowest number of locations with emerging malaria incidence conditioned on urbanization and basic education were detected in the bottom right panel where basic education was high with low urbanization (more rural). The results of the nine micromaps for each multivariate relationship indicated geographical correlations and co-location among the three variables simultaneously where two conditioning risk factors were having co-occurrence effect on the incidence of malaria (Figure 7a-k in Appendix). Co-location of all of the covariates, even when both had low proportions (bottom-left micromap in the panels) or where both had high proportions (top-right micromap) filtered specific districts affected by overall high malaria incidence.

The incidence of malaria increased markedly between 2010 and 2014. Our findings show that malaria remained a prevalent public health threat in Ghana with increasing trend among the districts despite numerous interventions in the country as reported previously [21, 27]. The increasing trend in malaria incidence might also have been induced by increased access to health care, especially provision of more community health planning services and clinics and free treatment of malaria at health facilities through National Health Insurance Scheme (NHIS) which encouraged more people to visit health facilities for healthcare during the study period. Significantly clustering of rates was consistently high in the uppermost parts of the country but also shifted gradually through the middle to the southern belts. Thus as steps were implemented to control the risk at one location heterogeneously, risks moved to low endemic areas, possibly due to human and vector migrations. The implication of our findings is that short-term impacts on transmission intensity, following scaled up insecticide-treated net coverage will be more difficult to achieve in high transmission settings like Ghana compared to areas where the majority of the population have a lower intensity transmission profile [28]. It is well documented that malaria transmission intensity exhibits strong spatial heterogeneity even at a local level in highly endemic areas [12‐14]. Despite few areas can be classified as relatively low risk areas (colspots), malaria incidence was generally high across the country with some districts experiencing relatively elevated burdens (hotspots). This finding is consistent with the climatic and environmental malaria risk factors analysis by Kumi-Boateng et al [29] and NMCP (28) in Ghana. Spatiotemporal clustering was relatively most observable in districts in the upper east region of the country.

We found a substantial number of the sociodemographic risk factors having significant influence on the transmission dynamics of malaria in Ghana. The weak correlations of the sociodemographic risk factors with the incidence of malaria indicate that these factors might not be major predictors of the occurrence of malaria. The contribution of sociodemographics, such as the employment-to-population ratio, agricultural activities, might be explained by several inter-connected factors. The poor would generally find it very difficult to afford the necessary prevention and control interventions such as anti-malarial chemotherapy, mosquito coil and repellents, access to clean and mosquito-free breeding environments and good housing units. Additionally, populations with low socioeconomic status reside in rural areas and often engaged in farming activities which increase exposure to the vector, have relatively poorer education and knowledge about malaria prevention, are more marginalised, and have less access to quality healthcare for prompt diagnosis and treatment. This also affects the health seeking and treatment behaviour in poorer rural areas [30] that can serve as reservoirs of the Plasmodium sp. parasite for the vector transmission. Housing the undiagnosed and untreated malaria parasites could facilitate the transmission through intra and inter-migrations and high mobility of the mosquito vector into low endemic areas or areas with lower levels of immunity [13]. Hence low socioeconomic status, limited education, low levels of urbanization, high migration and agricultural levels associated with greater levels of malaria morbidity across the local districts. Thus successes for malaria control also depend significantly upon knowledge, sociodemographic and socioeconomic status of the affected populations in endemic countries [15, 28] in addition to the biologic, climatic and environmental factors. In hyperendemic settings of malaria, the disease tends to spatially cluster based on different levels of environmental and climatic conditions and may also cluster with the sociodemographic factors [26]. One unexpected finding which is not well documented in literature is the influence of non-affiliation with religious groupings. Although the three main religious stratifications in Ghana did not correlate significantly with the incidence of malaria, non-religious affiliations significantly correlated inversely. The religious affiliations where people often come together and even stay overnight for prayers and other religious activities and at times in open spaces and some surrounded by bushes and in forest will expose people to the vector and those who might be having asymptomatic/ undiagnosed and untreated malaria. This means that targeted risk management strategies are needed for religious (Christianity, African traditional and Islamic) gatherings.

The ERMs and CCMs identified specific locations of high incidence that could be targeted [31] where malaria and specific sociodemographic factors were spatially interacting significantly. It is plausible that some of these sociodemographic factors could inform the management of the incidence of malaria in the identified districts. For instance, spatial interaction of malaria-basic education specifically highlighted districts in the middle and northern parts where increasing basic education and/or intensifying education campaign could contribute to reducing the expected incidence of malaria. Improving economic conditions in those districts (which will impact migration) could improve health seeking behaviour of the indigenes to help lower malaria incidence risk [13] in those districts. The ERMs revealed specific districts requiring extra attention that could be targeted for modifying a particular sociodemographic covariate in an attempt to lower the incidence of malaria, mostly at the northern and south-western parts. The conditioned micromaps of malaria located areas that are most/least affected of malaria with low and/or high accumulation of two risk factors. Our findings imply that although multiple factors are associated with malaria morbidity, none can be isolated as a sole target for malaria control. However, the ERMs and the CCMs highlighted specific locations that can be targeted for particular or pair-wise combinations of sociodemographic factors to address the spatial heterogeneous transmission of malaria. Due to the complex interactions observed in the risk factors and disease-risk factor spatial interactions, targeting each or pair could lead to extended or multiple control effects on malaria transmission. Specifically, improving employment-to-population ratio in rural areas will increase education attainment for improve knowledge and healthy lifestyle, increase the ability to afford good healthcare and good housing facilities, reduce rural-urban migration which in turn reduces urban slums with its associated health implications. These chain effects will significantly impact the successes for malaria control, especially in endemic areas.