Figure 2

Figure 2. Distribution of districts where epidemic meningococcal meningitis are likely to occur. a) observed distribution of meningitis epidemics (1841–1999). b) fitted model of observed meningitis epidemics based on environmental variables.

The earliest documented meningitis outbreak in Africa affected a French garrison in 1841 in Algiers (11), and >425 epidemics were documented at the subnational level for the next 158 years. These epidemics affected at least 1,231 (38%) of the 3,281 continental districts. Supporting our earlier findings, epidemics were not evenly distributed across the continent, instead affecting mainly districts in the Sahel and south of this region and extending from northern Uganda and the eastern part of Democratic Republic of Congo, through the Great Lakes and the Rift Valley regions to Malawi and northeastern Mozambique, and from northeastern Mozambique west and south to include other parts of southern Africa (Figure 2a) (1).

Absolute humidity, dust and rainfall profiles, land-cover type, and population densities were independently associated with the location of epidemics. However, we found that absolute humidity profile and land-cover type were the best predictors in the final multivariate model. Certain dust profiles and population density made only marginal difference to the performance of the model and were excluded to maintain simplicity. In addition, absolute humidity and rainfall profiles predicted similar risks in geographic locations in north and west Africa, but a model including the former performed better for the entire continent than one including rainfall; for this reason, we kept absolute humidity as a variable. Weighting for district size did not improve the performance of the model and was discarded.

The model, which is based on the absolute humidity profile and land-cover type, is described in terms of its baseline characteristics, estimated coefficients, standard errors, and contribution of the variables in Table 2. The risk map for epidemic experience in Africa derived from this model is presented in Figure 2b. The most important factor associated with the distribution of epidemics was humidity. Areas without a marked distinction between wet and dry seasons were less likely to have had epidemics than those with contrasting seasons. The areas without distinction between wet and dry seasons include deserts and the humid parts of coastal and central Africa, much of which are forested, and the areas with contrasting seasons comprise the semiarid savannah and grasslands of the Sahel and east and southern Africa. Surface maps of Africa demonstrated a close correspondence between humidity and land-cover types in these regions. The model also showed that, having accounted for the effects of humidity, sparsely vegetated and barren regions, areas of woodland mosaic, and shrub land were less likely than other regions to have ever had an epidemic. The Sahel, which has a prolonged dry season with low humidity was identified as the area with the greatest risk (p>0.6). Peripheral regions along its southern borders, where the dry season is shorter and less extreme, carry a moderate risk (p>0.4). The peripheral region extends from southern Sudan and Ethiopia to the Great Lakes and Rift Valley regions and parts of southern Africa peripheral to desert areas.

Figure 3

Figure 3. Receiver-operator characteristics curve for the model.

The ROC curve used to describe the performance of the model in terms of its sensitivity and specificity at various cutoffs; its overall accuracy independent of a single probability cutoff is shown in Figure 3. The model can discriminate between districts that have experienced epidemics and those that have never been affected. When the values for randomly selected districts were entered into the model, the epidemic risk assigned was higher for districts with epidemics than for those without in 82% of the cases. The areas identified by the model coincide to a large extent with the areas observed to have experienced epidemics. By using a probability cutoff value of >0.4 for predicting epidemic experience, the model had a sensitivity and specificity of 83% and 67%, respectively; these statistics were confirmed in the validation process (Table 3). The map of residual risk unaccounted for by the model had a random distribution, as expected for a model that worked well across the continent (not shown). According to the model, 7% of the population of Africa live in very high-risk areas, 17% in high, and 27% in moderate (based on 1990 estimates), corresponding to 44, 102, and 162 million people, respectively (Table 4).

This study represents a comprehensive and detailed description of the spatial distribution of meningitis epidemics at the district level in Africa and the first attempt to develop a spatial forecasting model for meningitis epidemics on the basis of the environmental characteristics of the continent considered a priori to be related to the spatial distribution of epidemics. The data have limitations that need to be considered for the proper interpretation of the models. For example, census data for Africa have limited accuracy (12), variables derived through remote sensing may only partially capture surface conditions, and data from meteorologic stations in Africa are often incomplete (13–15). The epidemiologic data span more than a century whereas the environmental and population data are relatively recent. In addition, some epidemics were likely never reported, and small outbreaks and those occurring in the 19th and early 20th centuries are likely to be disproportionately underrepresented. Problems in defining epidemics exist as well, since most reports lack clear or internationally recognized criteria; we had to accept the perception of an increased incidence that prompted outbreak reports. Moreover, the aggregation of local level statistics to an often larger and somewhat arbitrary district level, discrepancies between where people become ill and the location of notifying health facilities, and changes to district boundaries over time may each have resulted in potential loss of specificity. Despite these limitations, major outbreaks were unlikely to have gone completely unreported, and the long-term cumulative distribution of events is likely not misrepresented on a pancontinental scale. While population densities in Africa have increased greatly during the last 150 years and substantial land-use change (particularly in West Africa) is known to have occurred, the model was still able to identify the meningitis belt and areas previously described at risk beyond the Sahel (2–5,16); reports of epidemics occurring since 1999 have coincided with this description (available from: URL: http://www.who.int/disease-outbreak-news/).

The model also has its limitations. We are aware of those imposed on the analysis by spatial relationships in the data, which indicate that the importance of regression effects may be overstated (17). No perfect model exists, but one can be developed that is based on the simplest approach and combination of variables that can be used to distinguish between areas with high and low risk of epidemic experience. We were restricted by the limited availability of suitable datasets and chose the simplest combination of variables that best predicted outcome. While this choice is likely to have oversimplified the association between meningitis epidemics and the environment, the model has the advantage of using data that are available in the public domain and, being based on a simple combination of variables, are an important basis on which to develop research and future operational applications in resource-limited settings.

The model indicates not only that absolute humidity profiles and land-cover types can be used to distinguish between areas with high and low risk of epidemics but also that population density and dust may also be implicated. The incidence of meningococcal disease has previously been correlated with dry and dusty conditions in tropical and temperate climates (18–21). Humidity and land cover were included in the final model for statistical reasons, but dust and population density still have an independent effect and may be important in determining epidemic occurrence (22). The potential role of dust in precipitating epidemics is particularly interesting since dustiness in the meningitis belt has increased dramatically since the Sahelian droughts of the 1970s and 1980s. However, how environmental variables interact is unclear and remains the subject of extensive climatologic research. Furthermore, we did not take into account the effect of other nonenvironmental factors likely to be related to epidemics, such as population movement, vaccination coverage, and recent epidemics in the area. A combination of conditions is likely to be necessary for an epidemic to occur, and these nonenvironmental variables are likely to have additional predictive potential and should be considered in further studies.

Risk maps of vector-borne diseases in Africa based on environmental data have received considerable attention in recent years and are tools with public health potential (10,15). The mechanisms by which environmental factors influence meningitis epidemics in Africa are unclear (3). Areas within the traditional meningitis belt and beyond, however, are environmentally susceptible to epidemics with potentially large populations at risk; markers such as absolute humidity and dust profiles, land-cover type, and population density are independent predictors of these areas. In addition, rainfall and dust are predictors in some, but not all regions, and the potential to develop region-specific models that could be more sensitive within given ecologic zones warrants further study. Our findings could facilitate the development of models to identify regions with increased vulnerability to epidemics in the future and provide a basis for monitoring the impact of climate variability and environmental change on epidemic occurrence in Africa.