Using Rainfall and Temperature Data in the Evaluation of National Malaria Control Programs in Africa

Rationale for using climate information in malaria evaluations.

Central to any health impact assessment in an observational study is the concept of a baseline year or baseline period against which changes in outcomes can be measured. If the outcome is climate-sensitive we might expect that climate of the baseline year(s) chosen for evaluation purposes may have a substantial impact on the outcome of the evaluation. If malaria control intervention scale-up follows an unusually wet and warm baseline period and malaria incidence declines following interventions (during a drier and or cooler period), it may be tempting to attribute all of the decline in malaria outcomes to the investments in malaria control. However, correct attribution is important. As climate varies naturally over time, it is likely that the situation will at some point reverse (at least in part), resulting in an increase in climate suitability for transmission risk. If climate is not accounted for then higher malaria cases observed may be inappropriately attributed to program failure or other factors. Increasingly, climate data are being sought to remove this effect.^12–16

Below are five observations, given as examples, which indicate that climate variability and trends potentially impact the attribution of reductions in malaria morbidity and mortality solely to interventions.

1. Declines in vector populations (An. gambiae s.str and Anopheles funestus) have occurred independently of any known specific malaria intervention—for example, in Tanzania.¹⁷

2. Replacement of An. gambiae s.str with the more zoophilic and drought tolerant Anopheles arabiensis.¹⁸ For example, An. gambiae s.str. adult females from indoor collections predominated in Kisumu, Kenya, from 1970 to 1998 (ca. 85%). Beginning in 1999, when significant regional droughts occurred in eastern Africa,¹⁹ An. gambiae s.str. decreased proportionately relative to An. arabiensis, then precipitously declined to rarity coincident with increased bed net ownership as national bed net distribution programs commenced in 2004 and 2006.²⁰

3. Declines in malaria morbidity and mortality have preceded interventions. These declines have occurred on Pemba Island, Tanzania,²¹ and in a diverse range of ecological settings in Africa.²²

4. Declines in malaria morbidity and mortality have been shown to be associated with drought in, for example, Sudan and Eritrea^23,24 and across the Sahel.²⁵

5. Variations in monthly rainfall in rural Tanzania may be strongly associated with all-cause mortality (largely associated with malaria) with younger age groups and the elderly population more at risk.²⁶

Climate data for use in analysis.

The relative importance of climate to malaria transmission varies in different regions of Africa with semi-arid regions and upland/highland regions being most susceptible to climate-related impacts. Access to high-quality climate data at appropriate spatial and temporal resolutions and with national coverage and local relevance is key to the use of climate data in malaria impact assessment.

There are four types of climate data of potential use; 1) ground-based observations from meteorological stations—either as a point measurement or as a gridded surface, 2) climate proxies derived from satellite data, 3) proxies derived from combining observations with the output of global or regional climate models—for example, reanalysis products, and 4) climate indices, which represent regional or large-scale atmospheric or oceanic phenomena. Table 1 summarizes common data sources, their characteristics, relevance, strengths, and limitations for use in malaria impact assessment. Many of these data sources may also be accessed through the IRI Data Library^†; a powerful tool that serves over 400 public domain climate-related datasets.²⁷ High-resolution environmental data (e.g., Landsat) and satellite-based global climate and environmental data may be accessed through Google Earth Engine (https://earthengine.google.com). However, access to climate data is not sufficient by itself. Understanding the characteristics and quality of the data is critical to the proper integration of climate and environmental information in malaria studies.

It should be noted that the quality of globally available blended satellite, reanalysis, and ground station products depends critically on the quality, number, and distribution of the ground-based observations used in their calibration and creation. Access to station data is limited by the policy, cost recovery issues, and technical constraints of the National Meteorological and Hydrological Services (NMHS).²⁸ Unfortunately, few countries in Africa have quality-assured climate information readily available and accessible at an appropriate spatial and temporal scale and so there is, by necessity, considerable reliance on often poorly calibrated and unvalidated global products (e.g., from satellite data or reanalysis).

Integrating climate into specific malaria analyses.

Climate and environmental information have been used in a number of studies to help assess the impact of malaria interventions. Analyses range from a simple assessment of the climate pre- and post-intervention^16,29 to using climate predictors with malaria mortality,³⁰ cases,²³ or slide positivity³¹ as the predicted outcome. Incorporating both temperature and rainfall along with other environmental or socioeconomic variables into geospatial analyses using a Bayesian statistical modeling approach with malaria as the dependent variable is increasingly seen as the way forward for modeling spatiotemporal disease risk.^32–36 The use of climate-driven mathematical malaria models is another approach.^37–39 However, to-date, there is insufficient evidence that they can be used in malaria impact assessments at relevant spatial and temporal scales such as district level monthly analysis across multiple epidemiological zones (epidemic, endemic, etc.).

Herein, we present an approach that can be used when high-quality, long-time series malaria data is lacking but high-quality climate data such as rainfall and temperature are available.⁴⁰ Using specific operational tools that can incorporate climate data the approach is based on the general assumption that a warmer, wetter climate will likely increase the suitability for malaria transmission risk, whereas a drier, cooler climate will reduce it. In the absence of interventions or other social changes, we expect this to have a significant impact on malaria prevalence in areas where certain months of the year are deemed unsuitable for malaria transmission because it is too dry or too cool. We then assess the impact of choice of baseline years against which to compare intervention impact.

Analysis of climate suitability for malaria transmission during pre- and post-intervention periods.

To overcome the constraints identified for quality control and relevance of global climate products, the International Research Institute for Climate and Society (IRI), with support of the President's Malaria Initiative and other partners, has pioneered a new approach referred to as “Enhancing National Climate Services (ENACTS).”^28,41,42 Through this initiative, the availability of climate data is improved by careful quality control of data from the national observation network and combining all relevant station observations with satellite estimates for rainfall, and digital elevation models and reanalysis products for temperature. The new, quality assessed, spatially, and temporally complete (> 30 years for rainfall and > 50 years for temperature of 10-day data at 4–5 km²) ENACTS data products created allow for characterization of climate risks at a local scale with national coverage.

Developed rainfall and temperature (minimum and maximum) data and derived products are disseminated on the meteorological agency website^27,43 and through other means (i.e., on request via e-mail). The ENACTS process provides a substantive improvement on the number and quality of ground observations incorporated into blended historical and monitoring rainfall estimates and temperature products currently available (Figure 1).

In our study, monthly rainfall and temperature (minimum and maximum) data at 4–5 km resolution were obtained from the National Meteorological and Hydrological Services through the ENACTS initiative in Tanzania, Zanzibar, Rwanda, Ethiopia, and Mali.²⁸ Representations of the data—for example, climatology, anomalies, trends, and derived products are available online in “Maprooms” on the meteorological agency websites. Figure 2A–C presents an example of long-term anomaly time series produced for rainfall and temperature by the Ethiopian ENACTS Climate Analysis tool. An extended drought period from 1997 to 2009 for the Belg season can be observed. The year 2010 was extremely wet in both seasons and also had high minimum temperatures (compared with the long-term average). Substantial warming trends in minimum and maximum temperature can be observed in both seasons over the period 1983–2010. Similar analysis are readily undertaken using Maprooms in all countries where ENACTS has been implemented (Ethiopia, Tanzania, Zanzibar, Madagascar, Rwanda, Ghana, Zambia, and Mali). In addition AGRHYMET—based in Niger (http://cradata.agrhymet.ne/maproom/) has developed a regional ENACTS rainfall product for the Sahel, which incorporates approximately 500 operational weather stations from the participating countries.

Malawi, Angola, Uganda, and Mozambique are all countries where ENACTS products and services are absent or in development. Senegal had regional ENACTS rainfall data and this was compared with global rainfall products. For these countries, where quality rainfall and temperature information for use at local level is difficult to access, national level climate analysis were undertaken using rainfall derived from the Merged Analysis of Precipitation from the Climate Prediction Center (CMAP); a global product that is freely available. Previous work has shown that in Botswana, CMAP provides a reasonable approximation of observed rainfall at the national and seasonal time scale.⁹ We compared the results with Climate Hazard Group Infrared Precipitation (CHIRPS) rainfall estimate that involves satellite data merged with archived station data that has recently been made available via the U. S. Geological Survey.⁴⁴ This product is available for the entire globe at 5 km resolution but, as with other satellite-based data sources, the quality varies by country based on differing access to ground observations from the national meteorological and hydrological services. Both global products were tested against ENACTS data products, which are considered “gold standard” as they include all relevant climate data managed by the NMHS’s.

In addition, we explored options for accessing time series of minimum and maximum temperature that could be used to compare the Climate Suitability for Malaria Transmission (CSMT) pre- and post-intervention. Vancutsom and others⁴⁵ explored the possibility of retrieving high-resolution minimum air temperature data from the Moderate Resolution Imaging Spectroradiometer (MODIS) over different ecosystems in Africa. Comparisons between night MODIS Land Surface Temperature (LST) data with minimum air temperature showed that MODIS night-time products provide a good estimation of the seasonal pattern of minimum air temperature over different ecosystems; however, year-to-year variability was not explored. They did show that air time maximum temperature LST was very poorly related to seasonal patterns in maximum air temperature. Because of these limitations and because MODIS LST are only available since 2000 and so are not suitable for most of the analysis considered in this manuscript (as preintervention periods often included time before 2000) they were not considered further. To assess the historical climatology of temperature in different African countries, we used the University of East Anglia monthly gridded average air temperature surfaces climatology (1951–2000).⁴⁶ However, this data source was not suitable for the periods of our analysis; so instead, we accessed minimum and maximum temperature from reanalysis data obtained from National Aeronautics and Space Administration Goddard Space Flight Center MEERA^‡ after exploratory analysis indicated it was a good proxy at large spatial scales. We were able to access and analyze the global products through the IRI Data Library.

Specific tools for use by the malaria control community that are currently delivered through the ENACTS websites (http://iri.columbia.edu/resources/enacts/) in Tanzania (including Zanzibar), Rwanda, Ethiopia, Rwanda, and Mali include

1. CSMT tool⁴⁶ and

2. Climate Analysis tool.²⁸

3. Weighted Anomaly Standardized Precipitation (WASP) tool.⁴⁷

The CSMT tool was developed originally as a decision tool to support the timing of activities associated with malaria interventions in Africa.⁴⁸ It is an interactive map that displays for each month the percent occurrence of years when climate conditions were suitable for malaria based on the ∼30-year historic ENACTS data. Climatic conditions are considered to be suitable for transmission when the monthly precipitation accumulation is at least 80 mm, the monthly mean temperature is between 18°C and 32°C (P. falciparum) and the monthly relative humidity is at least 60%.^49,50 A modified tool, which uses a 16°C threshold has been developed to represent climate suitability for P. vivax malaria transmission. It is a useful tool for identifying where and when rainfall or temperature is likely a constraint to malaria transmission.

The Climate Analysis tool presents the history of rainfall and temperature data for any location or region as a time series of anomalies. This is particularly useful for temperature analysis where unusual increases in minimum temperature may increase the likelihood of malaria in highland regions identified as sensitive to temperature using the CSMT tool.

The original purpose of the WASP tool was to provide a simple visual means of relating country-averaged rainfall to a reference period of interest to explore the spatial and temporal patters of meteorological drought.⁴⁷ To compute the WASP index, monthly precipitation departures from the long-term average are obtained and then standardized by dividing by the standard deviation of monthly precipitation. The standardized monthly anomalies are then weighted by multiplying by the fraction of the average annual precipitation for the given month. These weighted anomalies are then summed over a 12-month period in this case, and this result is itself standardized.

WASP index values for use in malaria analyses were originally developed using rainfall data from CMAP⁴⁷ (https://iridl.ldeo.columbia.edu/maproom/Health/Regional/Africa/Malaria/MDG/index.html). We made a similar tool to analyze CHIRPS data at the national level and the tool was adapted for national/subnational use when high-quality ENACTS data became available. The latter can be found on the ENACTS Maproom sites at the participating NMHSs (iri.columbia.edu/resources/enacts).

For countries without ENACTS, we used CMAP and CHIRPS rainfall and MEERA reanalysis temperature accessed and analyzed through the IRI Data Library. In addition, we investigated climate trends and anomalies for Senegal using regional ENACTS data and global products. Using the CSMT tool from the IRI Data Library.²⁷ We identified Malawi, Uganda, and Angola as countries where warming temperature may play a role in favoring malaria transmission were further explored using reanalysis temperature data (minimum and maximum).

Comparison of different rainfall products for use at the national scale.

A comparable WASP analysis using the globally available CMAP and CHIRPS and ENACTS products rainfall products for Tanzania was undertaken (analysis not shown). Although showing clear similarities with the other products, CMAP rainfall (which includes limited station data from the Global Telecommunications System), did not capture well extreme rainfall including that of 1997/1998. Tanzania CHIRPS data appear very similar to the ENACTS national data when analyzed at the national level despite the fact that this product accesses fewer ground observations. Differences were observed at the subnational level (not shown). Based on these results, we used CHIRPS data whenever national ENACTS rainfall data were not available.

Of the 10 countries included in this study, six countries/regions (Tanzania, Zanzibar, Ethiopia, Rwanda, and Mali) had access to high-resolution ENACTS data for both rainfall and temperature. Regional ENACTS data were available for Senegal for rainfall alone and so both ENACTS and global products were used there. Analysis for an additional four countries (Malawi, Uganda, Mozambique, and Angola) was dependent on global rainfall and temperature products only.

Analysis of climate suitability for malaria transmission pre- and post-intervention.

Tanzania: the ENACTS-based WASP rainfall analysis for Tanzania using a baseline period of 1995–1999 and associated interventions for the 2000–2010 is presented in Figure 3. The baseline period (1995–1999) used in the analysis by Smithson and others¹⁵ includes rainfall associated with the 1997/1998 El Niño and the early part of the intervention period 2000–2005 includes three major droughts, making it highly plausible that the impact of interventions will be overestimated. Only a few highland regions of Tanzania are likely to be impacted by the observed warming in minimum and maximum temperature that can be seen in the ENACTS temperature products.²⁹

However, interventions in Tanzania were substantially increased post-2005, and therefore an evaluation comparing 1995–1999 and 2006–2010 may be more appropriate. If this new intervention period is chosen, then the intervention years are less likely to be influenced by the choice of an unusually wet (1995–1999) baseline period. Evidence of successful control against a more difficult climate increases the plausibility² that the interventions are having a significant effect.

Similar analysis has been done for other countries and regions using both ENACTS and CHIRPS data as available (Figures 4A–E and 5A–E).