GoKawiilMalaria prevalence data
We used a recently published database of P. falciparum prevalence in sub-Saharan Africa2. This compendium, compiled by Snow et al. over more than two decades, is one of the most spatially and temporally complete publicly available databases of infectious disease burden. The database covers the period from 1900 to 2016, although sampling has increased substantially since the turn of the century (pre-2000: n = 32,533; post-2000: n = 17,892). Most prevalence surveys used microscopy for diagnostics (n = 36,805) but a substantial portion of data also derive from rapid diagnostic tests (n = 11,154). The data have been compiled from a mix of archival research through public health documents, including the records of colonial governments and elimination campaigns from different periods; national survey data; electronic records published in peer-reviewed journals and grey data sources (for example, World Health Organization technical documents); and a mix of other sources compiled by international organizations. Records were georeferenced in the original study using a standard set of protocols, with a 5-km grid uncertainty threshold for point data, and broader areas stored as administrative polygons. In total, the data include a total of 50,425 prevalence surveys at a total of 36,966 unique georeferenced locations.
The Snow et al. data cover all available prevalence surveys, including all age ranges, but were converted by the authors of the original study to a standardized estimate of prevalence in children 2–10 years of age (PfPR 2−10 ), using a catalytic conversion Muench model. We chose to use these standardized estimates of childhood malaria prevalence because falciparum malaria has the highest mortality in children and pregnant women. The trends that we infer should generally be representative of broader transmission across age groups. In some cases, we note that declines in early-life exposure can lead to increases in incidence in adults61; however, these impacts are likely to be small, particularly given that active and passive improvements in malaria prevention, control and treatment much more directly determine trends in adult malaria risk.
Climate data
We used two sets of climate data in this study. The first is an observational dataset from the Climatic Research Unit (CRU-TS; v4.03 for model training and bias correction), which is constructed from monthly observations from extensive networks of meteorological stations from around the globe62. CRU-TS provides land-only climatic variables at a spatial resolution of 0.5° × 0.5° extending from 1901 to present (although our analysis is limited to the period 1901–2016). The second set of data is from ten global climate models (GCMs) selected from the sixth phase of the Coupled Model Intercomparison Project (CMIP6): ACCESS-CM2, ACCESS-ESM1-5, BCC-CSM2-MR, CanESM5, FGOALS-g3, GFDL-ESM4, IPSL-CM6A-LR, MIROC6, MRI-ESM2-0 and NorESM2-LM. In our historical analysis, we analysed (per GCM) one model realization of the ‘historical’ simulation, which includes anthropogenic greenhouse gas emissions, and one realization from the ‘historical-natural’ simulation, which includes only solar and volcanic climate forcing. For both the historical and historical-natural (hereafter and in the main text, ‘historical climate’ and ‘historical counterfactual’, respectively) simulations, we analysed the period 1901–2014.
To investigate the continued effect of climate change on malaria prevalence between 2015 and 2100, we analysed three CMIP6 future climate change simulations from each of the 10 GCMs. SSPs refer to the level of potential future global development (social, economic and technological) and the implication for climate change mitigation and/or adaptation actions or policy63,64. SSPs are combined with various possible future radiative forcings (RCPs) to form the climate change scenarios used in CMIP6. Of the available SSP–RCP scenarios, we selected and used three. The first two suggest enhanced human development outcomes with increased potential towards a more sustainable (SSP1)65 or a less sustainable (SSP5)66 economy. The third, SSP2 (ref. 67), is a mid-way scenario, which assumes a future that mostly follows historical trends64. We selected these scenarios in combination with a low (SSP1–RCP2.6), intermediate (SSP2–RCP4.5) and high (SSP5–RCP8.5) greenhouse gas concentration scenario.
We applied a standard quantile–quantile (Q–Q) bias-correction68,69 to the CMIP6 precipitation and temperature datasets for both of the historical simulations for the period 1901–2014, and all three future simulations for the period 2015–2100. Before the bias correction, we first remapped all simulated CMIP6 precipitation and temperature datasets to the same grid cell size (0.5° × 0.5°) as the CRU-TS observation data. We then performed for each CMIP6 model, the Q–Q bias correction at each grid point by mapping the quantile values (q i ) for the empirical cumulative distribution functions for each of the 12 months over the period 1901–2014 (for each grid point) onto the corresponding quantiles in the observational dataset (CRU-TS), so that the observed precipitation or temperature values associated with q i become the bias-corrected value in the simulations. For the counterfactual (and future) simulations, we first determined, at each grid point, for each value of precipitation or temperature (for each month) over the period 1901–2014 (2015–2100), the equivalent quantile (q j ) in the factual simulation and then identified the precipitation or temperature value associated with q j in the observational dataset as the bias-corrected value. We detrended both precipitation and temperature datasets before applying the bias-correction procedure, and then added the trends back after69.
Spatial data aggregation
Our statistical analysis is designed to isolate variation in the weather that is uncorrelated with other socioeconomic and/or environmental factors that influence malaria prevalence. As detailed in the next section, we build on a large body of climate econometrics research38,40,70 to do so, estimating a model that leverages variation over time in weather conditions within the same location. To estimate such a model, we required observations of malaria prevalence covering the same region in multiple time periods. By contrast, the raw prevalence data that we obtained from ref. 2 are point data observations from individual surveys conducted at different times, such that single geolocations are not observed repeatedly over time. Therefore, we aggregated the point-level data from ref. 2 by averaging PfPR 2−10 observations to the first administrative level within each country (that is, state or province level, or as shorthand, ADM1), using shapefiles provided by the Database of Global Administrative Areas dataset v3.6 (www.gadm.org). This level of aggregation provides sufficient granularity to capture differences in climate impacts within countries and to control for local heterogeneity in confounders, while ensuring sufficient data coverage within these units. This aggregation scale has also been conducted in previous work that models this dataset at the same spatial resolution2. For robustness, we also show results from a statistical model that does not aggregate data, and instead uses the prevalence data at its native resolution (see below for details).
To compute average prevalence values at the scale of ADM1, we used an unweighted arithmetic mean over all prevalence surveys observed in the corresponding ADM1 month. This approach imposes minimal assumptions on the spatiotemporal process of malaria transmission and requires no additional high-resolution data (for example, population) for use as weights, which are unavailable for sub-Saharan Africa for years as early as 1901. Although previous work aiming to construct comprehensive high-resolution estimates of health outcomes using point data often uses spatiotemporal smoothing methods (for example, ref. 71), doing so here would artificially introduce spatial and temporal correlations that could bias recovered regression coefficients and threaten inference72.
... continue reading