Climate drivers of vector-borne diseases in Africa and their relevance to control programmes

Climate and vector borne disease

Many parasitic, viral, and bacterial diseases respond to variations in the climate whether through their geographic distribution, seasonality, inter-annual variability, or temporal and spatial trends. Detailed reviews of climate variables and the impact on pathogen and vector dynamics are available for a wide range of diseases [1, 2].

Known relationships of climate variability and change and the climate-sensitivity of most important infectious diseases causing considerable morbidity and mortality worldwide suggests the potential role of climate information in improving climate sensitive health outcomes [3]. Although many infectious diseases of humans are climate sensitive – those that are transmitted by arthropod (insect and tick) and snail vectors are particular important in lower and middle income countries [4]. They are therefore prioritized by the Tropical Disease Research [5] initiative of the World Health Organization and partners [5, 6].

The African climate system and its drivers at multiple time-scales

The health and wellbeing of African populations is closely tied to their environment which is itself closely linked to the regional and local climate. An extreme range of climates span the continent, according to the Köppen-Geiger classification system (Fig. 1) [17]. Across the continent the climate varies from arid zones (including the Sahara, Somali-Chalbi and Kalahari deserts), steppe or semi-arid regions (e.g. Sahelian savannah) to humid tropical environments (Congo river basin). Humid subtropical climates are features found predominantly in southern Africa but also include areas in the Ethiopian highlands. In some regions these widely diverse climates co-exist within relatively small areas and rainfall amount and seasonality (for example) may change significantly over tens of kilometers [18]. The changes in seasons (particularly the rainy and dry seasons) is the dominant characteristic of regional climate and it consequently drives the seasonal pattern of human activities as well as vector-borne diseases across the continent. The large seasonal variations in rainfall that distinguish different climate zones is seen clearly in Fig. 2a–d –which indicates the fraction of mean annual rainfall that falls within 3 month seasons (December–February: DJF; March–May: MAM; June–August: JJA; September–November: SON). The Fig. 2b and d indicate that East Africa has a bimodal season while others, such as the Sahel (see Fig. 2c) have a single rainy season, more typical of monsoon behavior.

The most significant driver of seasonal temperature change across Africa (where proximity to the equator might suggest nearly constant year-round temperatures) is the monsoonal rains, in part related to the inter-tropical convergence zone defined previously. For instance, cloud cover at night will tend to increase minimum temperatures whereas cloud cover in the day time will tend to reduce maximum temperatures [19]. These different responses indicate that minimum and maximum temperatures are better treated as separate variables rather than combined as mean temperature.

Whereas weather is almost entirely governed by conditions in the atmosphere, the climate is substantially driven by slower processes, particularly in the major oceans. The climate at any location varies from its mean historical climate state on multiple time-scales, from annual to multi-decadal (10–30 years) to long-term climate change; the latter compatible with anthropogenic climate change signals. The magnitude of these variations and trends may enhance or decrease the climate suitability for different disease vectors and their pathogens.

Sea surface temperature variations in the Atlantic [20], Indian [21] and Pacific [22] oceans influence the African climate on different time scales. We consider three timescales of variability in the African climate that describe the past and provide some indication of the future. El Niño-Southern Oscillation (ENSO) is the most important driver of climate variability at seasonal-to-interannual timescales [23], a key source of climate predictability in Africa [24] (see Fig. 3) and relevant to the development of climate information services targeting health decision-makers [3]. It is important to recognize that ENSO (El Niño and La Niña) impact the climate (and thereby climate-sensitive health outcomes): (a) differently according to the variable of interest (e.g. rainfall, and minimum and maximum temperature), (b) at different spatial scales, (c) in some regions and not others, (d) in some seasons and not others, (e) often according to its strength, and sometimes in a non-linear fashion, (f) at varying periods (from 5 months to ~ two years), with both El Niño and La Niña events on occasions occurring in the same calendar year (e.g., 2010), (g) often substantially conditioned on the action of other climate drivers, such as the Indian Ocean Dipole [25].

Natural variations in the climate at 10–30 year time frames (decadal) have also been observed in Western, Eastern and Southern Africa and again may be specific to region and season. In Eastern Africa decadal rainfall variations are largely confined to the long rains which occur between March and May [26]. Where historical data is sufficient, long term trends in temperature and rainfall, consistent with climate change, may be established once the noise from shorter term natural variations in the climate have been removed. Unless the impact of the different timescales can be disentangled, there is considerable opportunity for confusion, with important implications for decision-making and potential maladaptation. For instance, climate change models have indicated that Eastern Africa will become wetter towards the end of the twenty-first century while the region has, since 1999, experienced an increased frequency of drought [27].

Here we aim to characterize the African climate – its variability, trends and potential predictability – and establish the relevance of this knowledge and current tools to operational vector-borne disease control efforts.

International research Institute for Climate and Society (IRI) data library

Analysis 1. Multi-timescale climate decomposition

To better understand how much of the total variance in rainfall and temperature anomalies across the African continent is explained by different time-scales, a ‘timescale decomposition’ methodology [31] was used. The temporal analysis was focused on the WHO-TDR study sites. This approach has been used elsewhere to explore the contribution of climate variations and trends at multiple timescales to the observed seasonal climate of Latin America associated with the 2015 Zika virus epidemic [32].

Analysis 2: Assessing the impact of the ENSO on rainfall and temperature across Africa

In Africa ENSO impacts on African rainfall are well known and vary according to region and season [24]. While the impact of ENSO on global tropical temperatures is also widely appreciated [34], local effects are amplified or muted by ENSO impacts on rainfall [19]. The rainfall response to ENSO is nearly contemporaneous however, this may not be true for temperature. Once El Niño has begun, there is a ramp up of global temperatures which are then slow to dissipate after the return to a neutral phase although they may cool down rapidly if La Niña conditions emerge.

Analysis 3: Assessing the local impact of ENSO on rainfall and temperature in Tanzania

The analysis for one of the WHO-TDR study sites Monduli, Arusha, Tanzania – was further investigated using products and services from the ENACTS initiative [28]. ENACTS national climate products (rainfall and temperature) are created by quality – controlling all national station observations and combining this data with data from proxies – satellite estimates for rainfall, digital elevation models, and reanalysis products for temperature. The approach thus combines the spatial information from the proxies with the accuracy from point station measurements. The final products are datasets with 30 or more years of rainfall and temperature time-series data at a ten-daily (dekadal) time scale for a 4-km grid across the country. ENACTS products and services are disseminated online via Maprooms that are developed using the lRI Data Library which is installed at the Tanzanian Meteorological Agency [30] as well as in a number of other African countries (iri.columbia.edu/resources/ENACTS). This online mapping service provides user-friendly tools for the analysis, visualization, and download of climate information products via the NMHS websites.

Analysis 4: Assessment of seasonal rainfall and temperature predictability across Africa

Having identified the dominant signals of rainfall and temperature variability and trends in the different regions of the African continent, we explore their predictability using a two-tiered atmospheric global circulation model forecast system based on sea surface temperatures.

Analysis 1. Multi-timescale climate decomposition

The results of the timescale decomposition analysis for rainfall and temperature are presented in Figs. 4 and 5. Note that while the decomposition of a signal into trend, low- and high-frequency components may seem straightforward, the analysis presented involves a number of subtleties that are described in detail in documentation that can be found on the timescale decomposition Maproom site (see Table 1). The documentation also offers a number of caveats regarding the interpretation of Maproom displays.

Analysis 2. Assessing the impact of the ENSO on rainfall and temperature across Africa

The positive and negative impact of the El Niño on rainfall in the October to December for Eastern Africa and July to September seasons (for the Sahel) respectively are presented in Fig. 6a&b, while Fig. 6c indicates the positive impact of La Niña conditions on the rainfall of Southern Africa during the main season (December to February). On the other hand, Fig. 6d shows no impact of El Niño on the main rainy season (March to May) in Eastern Africa. Additional analyses for other seasons and for temperature can be obtained directly from the Maproom (Table 1).

The relationship of ENSO states to seasonal rainfall totals and average annual temperature time series are presented for Botswana in Fig. 7. The colour bars indicate the ENSO phase for an individual year, and the horizontal lines show the historical tercile limits. The image allows a quick assessment of the historical impact of ENSO by region and season and gives a visual indicator of the spread of results.

Note that the ENSO Maproom does not provide a forecast, but is a good tool for exploring the effect of different ENSO phases on seasonal rainfall and temperature. It is based on historical observations of rainfall and SST alone. Where a strong signal is found, it suggests that there is an opportunity for skillful seasonal forecasts since such forecasts substantially rely on a strong ENSO signal.

Analysis 3. Assessing the impact of the ENSO on rainfall and temperature in Tanzania

The results of the Pan-African ENSO analysis above were repeated in a national scale analysis using ENACTS products and services made available by the Tanzanian Meteorological Agency on their website (Table 1). The analysis indicates a moderate to strong impact of El Niño across the country associated with the Oct–Dec short rains (Fig. 8). A detailed analysis of the ENSO rainfall and temperature interaction for Monduli District, Tanzania (Fig. 9) for October–November-December (OND) is presented in Fig. 10a&b. Figure 10a indicates that El Niño years [41] have rainfall amounts predominantly within the normal to above normal range whereas La Niña years (blue) have rainfall amounts predominantly within the normal to below normal range. Figure 10b indicates that El Niño years [41] have minimum temperatures that are predominantly within the normal to above normal range whereas La Niña years (blue) have minimum temperatures predominantly within the normal to below normal range. Similar analysis indicting the correlation of the positive and negative phases of the Indian Ocean Dipole where completed using the Tanzanian Meteorological Agency (TMA) Maproom (not shown here). The same analysis can be done for Kenya and other ENACTS countries.

Analysis 4. Assessing the predictability of seasonal rainfall and temperature across Africa

The skill of seasonal climate forecast across Africa, as measure by the Generalized Relative Operating Characteristics (GROC) metric, for temperature and rainfall forecasts averaged over the entire year is poor (see Fig. 11a&b). However, both temperature and rainfall seasonal forecasts demonstrate skill in certain regions when particular seasons are considered. For example, during DJF, temperature forecasts tend to be good in southern Africa where they coincide with the main rainy season and also in parts of western Africa. They are also skillful in eastern Africa for both rainfall and temperature despite the short rainy season being largely confined to OND (see Fig. 11c). Rainfall in the Sahel exhibits some predictability during the main July–August-September (JAS) season. Although it is not very high, the skill of forecasts for rainfall for this season is on average higher than surface temperature skill (see Fig. 11e&f). Note that the crude nature of the climate data used in the analysis will limit the evidence of predictability.

Climate information into national decision-making for vector control purposes

Forecasting vector-borne diseases, such as malaria, using climate information is not new. Over a century ago records of unusual rainfall along with impoverished foods stocks were used as indicators of forthcoming malaria epidemics in the Punjab region of India [42]. In recent years an extensive research literature has emerged on the predictive relationship of observed and forecasted climate events in Africa and the spatial, seasonal, year to year and longer term shifts in vector-borne diseases [1]. Furthermore there has been an increase in studies providing evidence of the skillfulness of vector-borne disease forecasts based on climate monitoring products and seasonal climate predictions [14, 43, 44] and a greater interest in such analysis by policy-makers [4].

A key challenge to accurately using climate information for vector-borne disease prediction is the spatial and temporal variability in climate variables of interest. While a range of variables may be relevant to transmission they may not be available for use in operational systems which require national coverage, relevance at the local scale and near real time updates. Temperature and precipitation conditions may be predictable in one region or season but this does not necessarily mean that it can be extended to another. The series of analysis presented here are designed to establish which timeframes of variability are most important and reliable for disease prediction in the different study regions.

Analysis 1. Multi-timescale climate decomposition

The timescale decomposition analysis revealed that while century long term changes in rainfall were not a major historical concern across Sub-Saharan Africa during the twentieth century, decadal-scale variability has significant impacts on the climate, and hence populations and economies, in strongly affected areas such as the Sahel. This region shows the most extreme variations of seasonal climate anywhere in the world. Dramatic year to year variability in rainfall (in part related to ENSO events) is super-imposed upon decadal shifts in the climate as well as a long term drying trend. However, climate change models are uncertain as to the sign (wetter or drier), let alone the magnitude of potential changes in rainfall in this region. The decadal fluctuation in West African rainfall observed in Fig. 5b has been linked, in other studies, to SST variations in the Atlantic Ocean although the Indian Ocean may also be playing a role [20]. The long decline in rainfall during the 1970s and 1980s in the Sahel contributed to the retreat of malaria in this region [45]. The return to a higher rainfall regime in the last two decades (also likely a decadal variation) may have contributed to the re-emergence of Anopheles funestus to some areas, including Niger, after an absence of many years [46].

In East Africa, there has been a significant drying in the climate over the last two decades (Fig. 5c). This has occurred at a time when climate change models project that East Africa is getting wetter in the future – a phenomena called the “East African Climate Paradox” [22]. According to Lyon, the observed drying started abruptly in 1998 with a steep decline in the long rains (MAM) and is found to be driven strongly (although not necessarily exclusively) by natural decadal variability in the tropical Pacific rather than anthropogenic climate change [47]. The East African short rains (OND) are not affected by this decadal process further indicating distinct nature of these two seasons. As March–May is the main rainy season throughout much of Eastern Africa a dramatic decline in rainfall amounts in this season is likely to have a profound effect on vector-borne diseases such as malaria in affected areas [48].

There is also evidence of decadal variability in rainfall in Southern Africa (Fig. 5e) which has a tendency to become wetter during decadal periods when the eastern Pacific Ocean is cooler than average [47]. Mason and Jury [49] suggest there may be some periodicity of decadal variations in the climate of South Africa having a dominant period of about 18 years.

Continued warming of the planet is the most certain feature of climate change models [50]. Warming trends over the last century (and in particular from the 1970s, is evident in all regions of Africa where data is sufficient for analysis (see Figs. 4d and 5bd,f). For instance there is now substantive evidence that the East African highland region has been warming over the last 30 years [19, 51] with potential impacts on malaria and other vector-borne disease transmission in areas where transmission has hitherto been constrained by low temperatures.

Analysis 2: Assessing the impact of the ENSO on rainfall and temperature across Africa

Our results are consistent with what is known about ENSO and the climate of Eastern Africa. Here the annual cycle of rainfall tends to be bi-modal, with two physically and statistically uncorrelated rainy seasons [26] occurring in October–December (short rains) and March–May (long rains). Year-to-year variability of the short rains is frequently associated with ENSO [24]; but this connection depends critically on sea surface temperatures in the Indian Ocean, not just the Pacific. El Niño is typically associated with wetter than average conditions, while La Niña is frequently associated with drought in the short (OND) rainy season. A positive Indian Ocean Dipole (IOD) [52] is also associated with enhanced short rains; its opposite phase with drier than average conditions. While we have not undertaken an IOD analysis the relationship and can be explored in local East African ENACTS Maprooms (iri.columbia.edu/ENACTS).

Rainfall in many parts of the northwestern region of Eastern Africa (western Ethiopia and parts of western Kenya) have a boreal summer rainy season from June–September which is more in common with the timing of the Sahelian rainy season. The climate of the Sahel exhibits typical monsoon behavior, with a single peak in the rainy season between June–September. Our results support other studies which find a modest connection between ENSO and seasonal rainfall variability in the Sahel [53] with El Niño events associated with drier than average conditions and La Niña with wetter than average conditions.

Our results are also consistent with what we now about the climate of Southern Africa which is influenced by atmospheric circulations in both the tropics and the mid-latitudes. The main rainy season typically extends from October–April across much of the region, peaking during the southern-most extension of the inter-tropical convergence zone. By contrast, the southern tip of South Africa has a maximum in rainfall during the southern hemisphere winter season (May–September), associated primarily with the passage of mid-latitude storm systems [49]. A relationship between seasonal rainfall variability and ENSO has been observed in the region [54]. El Niño events are typically associated with drought in Southern Africa with La Niña linked to wetter than average conditions, although even strong El Niño events are not necessarily accompanied by drought [55]. There is substantive evidence that malaria in southern Africa is affected by SSTs in the Eastern Pacific (the Niño 3.4 region) with La Niña events frequently associated with an increased occurrence of cases [56, 57].

While we have not considered in detail the climate of Central Africa, we note that it contains the second largest area of tropical rainforest on earth and is therefore an important, but poorly studied, part of the global climate system [41]. It also has a high burden of malaria. The annual cycle of rainfall shows a bimodal behavior, with relative rainy seasons peaking in March–May and October–December, although there is substantial rainfall outside these seasons. The variability of the climate of Central Africa has received comparatively little attention compared to other parts of the continent [58]. On seasonal to inter-annual timescales, some studies have suggested a relationship between rainfall variability in Central Africa and SSTs in the tropical southern Atlantic Ocean [59]. For example, warmer than average SSTs off the Angolan coast are associated with increased rainfall, particularly in the March–May season and in the western part of the region. It should be noted that the quality of climate data for this region is extremely poor with few operational meteorological stations available. Consequently, global products for this region are likely also poor.

Our results (Fig. 5a, b) are consistent with other studies that show only a weak link between seasonal rainfall variability and ENSO in Central Africa with the largest connection found during the boreal fall season where El Niño (La Niña) events are associated with drier (wetter) than average conditions [60].

It is to be expected that the signal of the inter-annual relationship between climate and vector-borne diseases in Central Africa will also be weak as the environment is consistently warm and humid with high levels of rainfall throughout much of the year. Variations are likely insufficient to impact on transmission although there is scant vector or case data to establish whether or not this is the case.

The value of high resolution climate data in assessing the impact of ENSO on rainfall and temperature at the subnational level

National climate datasets made available through the Enhancing National Climate Services (ENACTS) initiative, provide additional insights into the relationship of ENSO (and the Indian Ocean Dipole) to rainfall and temperature variations at spatial scales which are relevant for vector-borne disease monitoring and prediction. The higher quality data sets are created from a blend of all the relevant observations made available by the National Meteorological and Hydrological Services, with the best global products. The improved quality of the data sets over global products make it easier to reveal the predictability that exists. Similar analysis are now possible in all countries where ENACTS is being implemented (see Table 1).

The relative importance of climate drivers and their potential predictability

The relative importance of the three categories of climate drivers and their predictability are region and variable specific. For year-to-year-variations, ENSO is the predominant driver of variability in rainfall and temperature and ENSO impacts on the climate can be observed most strongly during the single rainy seasons of Southern Africa and the Sahel and the short rains of Eastern Africa. Decadal variations in rainfall are also significant in the Sahel and have been observed for the March–April–May rainy season in Eastern Africa (not shown). Long term trends are observed the temperature data for southern and western Africa but the analysis for eastern Africa is constrained by data quality. Challenges encountered when seeking predictions at climate timescales are outlined in Table 2. In particular, our ability to assess forecast/prediction/scenario skill at different time scales is constrained by the lack of sufficiently long historical climate data. To observe the accuracy of a weather forecast one needs to wait a day or two and then the expired forecasts can be assessed against what is observed. Within a season there is plenty of data which can be used to assess forecast skill. For seasonal prediction, many regions only have one or at most two rainy seasons. Since seasons may act independently they each need to be treated in separate analysis. Thus assessing the skill of a probabilistic seasonal climate forecasts requires a minimum of 30 or more years of climate data against which the forecast models can be run in “hindcast mode”. Seasonal climate forecasts (both rainfall and temperature) are predicted shifts in the probability density function of seasonal rainfall totals or temperature means relative to a climatological baseline. The forecasts are commonly expressed in tercile probability format (i.e., probabilities of below-normal [BN], near-normal [NN] and above-normal [AN] rainfall or temperature categories). Thus, within a pdf of 30 years of climate data we have 10 years BN, 10 years NN and 10 years AN. With this short time series signals have to be very strong to be statistically significant. Describing a year as above-normal, provides little indication of the likely outcome in terms of disease. Is the season likely to be extremely wet? above a certain rainfall threshold? with rainfall events well distributed over time?. These types of questions are increasingly being addressed by climate scientists and we may expect much more nuanced seasonal forecasts to be available in the near future.

The quality of the data used to assess forecast skill also matters. If the data set gives a poor indication of actual climate conditions, then the skill test results are likely to be poor. Where available the ENACTS historical climate data (30+ years) provides a high quality climate data set for use as the forecast predictand.

The challenge of verifying forecasts that will happen decades into the future become even more onerous. As there are few places in the world where historical climate datasets go back sufficiently long in the past to assess variability over 10–30 year time frames, a more general validation of the model is needed. This is based on an understanding of its underlying mechanisms and the relationship of model outputs to historical climate characteristics of the region of interest. This is also true for the assessment of climate change model outputs.