The historical fingerprint and future impact of climate change on childhood malaria in Africa

A robust signal of climate sensitivity

Over the last century, the prevalence of childhood malaria has exhibited a strong concave relationship with temperature (Figure 2A). Closely aligning with theoretical expectations that P. falciparum transmission by Anopheles gambiae mosquitoes should peak around 25.6°C ⁶, observed values of Pf PR₂₋₁₀ in our dataset peak around a monthly mean temperature of 25.8°C (Figure S1). Based on these biological expectations, we estimate the effect of temperature as a quadratic response in a panel regression model, and find that prevalence peaks at 24.9°C (95% confidence interval across the 1,000 bootstrapped models: 22.5°C, 27.0°C). These results confirm that laboratory-based studies approximate malaria epidemiology in real populations quite well, and that temperature plays a substantial role in transmission dynamics: a 10°C increase or decrease from the optimal temperature lowers prevalence by ∼8 percentage points.

• Download figure
• Open in new tab

Figure S1: Malaria prevalence follows biological expectations.

(A) The theoretical expectation for R₀(T), the scaled partial response of the basic reproduction number to temperature, estimated based on laboratory experiments (black line) ⁶. Transmission peaks around an estimated optimum of 25.6 ^◦C (grey dashed line). (B) Observed malaria prevalence data from Snow et al. ⁴ matched to monthly temperature from CRU weather station data, summarized and smoothed using a generalized additive model. The observed optimum temperature (red dashed line) closely matches expectations based on laboratory experiments (grey dashed line). (C) Main panel regression estimate for prevalence response to temperature (also shown in Figure 2A). The modeled optimum temperature (red dashed line) is slightly lower than in laboratory experiments (grey dashed line). (D) Histogram of optimum temperatures derived from 1,000 bootstrapped estimates of the panel regression model shown in panel (C). The mean optimum temperature across all bootstrap samples (red dashed line) is identical to the optimum shown in panel C.

The relationship between precipitation and malaria prevalence is more complex ⁴⁹, and likely less consequential for historical trends (Figures 2B, 2C). Contemporaneous monthly precipitation exhibits a nonlinear, but highly uncertain, relationship to prevalence (see Figure S12). To parsimoniously capture nonlinear effects and disentangle divergent impacts of low and high precipitation, we define precipitation shocks with two binary indicator variables, equal to one when monthly precipitation falls below 10% (we label this a “drought shock”) or above 90% (“flood shock”) percentiles of monthly precipitation calculated for each subnational unit. While drought and flood events are complex phenomena, which develop from the combination of multiple factors (e.g., soil conditions and topography) in addition to rainfall over varying timescales, we use this drought/flood terminology as shorthand to indicate extremely low or high precipitation months. Although most effects are statistically insignificant, we find that drought shocks tend to decrease malaria prevalence 1-2 months later, while conversely, flood shocks have a positive effect on prevalence 2-3 months later. These effects and their timing are broadly consistent with expectations about how precipitation mediates the availability of mosquito breeding habitat: dry-out kills larvae and eggs, while inundation creates new breeding habitat ^{50,51,52,53,54,55,56,57}. Sensitivity analyses were also weakly suggestive of another established mechanism, in which floods may wash away eggs and larvae, reducing transmission in the shorter term (see Figures S10 and S12). Overall, extreme precipitation has a measurable effect on malaria prevalence, but may be less important than temperature; however, given the sparsity of weather station data⁵⁸ and the uncertainty of precipitation reconstructions⁵⁹, it is also possible that our analysis unavoidably underestimates the effect of precipitation due to measurement error.

Additional sensitivity analyses reinforce that these prevalence-climate relationships are both statistically robust and biologically consistent. Key findings are generally insensitive to alternative model specifications, such as: the inclusion of lagged effects of temperature (Figure S7); higher-order polynomial effects of temperature (Figure S11); alternative definitions of drought and flood shocks (Figures S8-S10); and alternative spatiotemporal controls, which account differently for variation over space (at region, country, and state levels), time (including yearly and monthly variation), and interactions among space and time (Figure S6 and Table S2).

Historical impacts of climate change (1901-2014)

We find that anthropogenic climate change has, more likely than not, been responsible for a small increase in the average prevalence of childhood malaria across sub-Saharan Africa since 1901 (Figures 2D). Compared to counterfactual simulations without anthropogenic climate forcing, we estimate that by 2010-2014, anthropogenic climate change had caused an increase in continental mean Pf PR₂₋₁₀ of 0.09 percentage points (p.p.; 95% confidence interval: -0.30 p.p., 0.51 p.p.). Simulations with an attributable increase in continent-wide mean prevalence outnumber those with losses by two to one (proportion P₊ of 10,000 paired factual versus counterfactual simulations with a positive difference in prevalence = 0.66). These increases are almost entirely driven by rising temperatures from anthropogenic climate forcing; the effects of drought and flood events on prevalence show no distinguishable signal from anthropogenic climate forcing over time (Figure S4). This overall trend masks substantial regional heterogeneity in historical climate change impacts (Figure 3A; Figure S2), driven almost entirely by elevational and latitudinal gradients in temperature (Figure 3B,C). For example, attributable changes in prevalence across southern Africa are high in both magnitude and certainty, with an overall increase of 0.63 p.p. (95% CI: -0.04 p.p., 1.40 p.p.; P₊ = 0.97)—nearly an order of magnitude greater than the continental mean. In contrast, climate change has contributed to significantly lower malaria prevalence in west Africa (mean = -0.38 p.p.; 95% CI: -0.90 p.p.,0.02 p.p.; P₊ = 0.03), where temperatures already often exceed the biological optimum for transmission. In the central African basin, a stronghold of malaria endemicity with average temperatures close to the 25^°C optimum, the change in prevalence attributable to anthropogenic climate change is positive, relatively small, and uncertain (mean = 0.18 p.p.; 95% CI: -0.19 p.p., 0.61 p.p.; P₊ = 0.83). Finally, we estimate a meaningful overall increase in prevalence attributable to anthropogenic climate change in east Africa (mean = 0.33 p.p.; 95% CI: -0.14 p.p., 0.87 p.p.; P₊ = 0.91), but note that changes in prevalence are distributed unevenly along the steep elevational gradient: increases of up to 1-2 p.p. in the Ethiopian highlands and the greater Rift Valley region are accompanied by small but significant local declines throughout lowland areas in Ethiopia, Sudan, South Sudan, Eritrea, and Djibouti.

• Download figure
• Open in new tab

Figure S2: Historical changes in malaria prevalence attributable to anthropogenic climate change from 1901 to 2014.

Map shows the estimated change in prevalence attributable to anthropogenic climate change in each administrative unit, based on the difference between the historical climate in 2010-2014 and a counterfactual scenario for the same period simulated without anthropogenic warming. Polygons with a black solid outline indicate areas with changes that were statistically significant (α = 0.05) based on the sign of 10,000 bootstrapped simulations. Mean estimates shown here provide the same information as in Figure 3, but on a single color scale (i.e., no uncertainty visualization).

While these effects are meaningful, we caution that they are also far smaller than the reduction achieved through healthcare, mosquito nets, vector control, and economic development: previous work with the same dataset has estimated a reduction since 1900 of 16 p.p. (that is, a continent-wide decline in average Pf PR₂₋₁₀ from 40% in 1900-1929 to 24% by 2010-2015⁴), while our estimates of historical climate change-attributable changes rarely exceed 1.5 p.p. for any individual administrative region. Additionally, we estimate that average reductions in prevalence realized during the Global Malaria Eradication Program (1955-1969; estimated reduction averaged over the entire period: -4.80 p.p.) and recent programs like Roll Back Malaria and the Global Technical Strategy (2000–2014; estimated reduction averaged over the entire period: -3.35 p.p.) were substantially larger than the cumulative effects of anthropogenic climate change (Table S2). Relatively small and spatially differentiated climate-related changes in burden could have been easily concealed by the greater impact of these programs, highlighting both the success of elimination programs and the importance of using an empirical approach like ours to isolate the effect of climate from other co-evolving factors.

Future impacts of climate change (2015-2100)

Despite contemporary trends, we project that within the next quarter-century, anthropogenic climate change will begin to reduce the prevalence of falciparum malaria in sub-Saharan Africa (Figure 2D; Table S1). This trend is driven largely by rising temperatures in lowland areas north of the equator, with greater possible reductions in higher greenhouse gas emissions scenarios (Figure 4). In these scenarios, temperature-related declines are slightly offset by floods, which will become more frequent across Africa ³⁶, although their impact on overall trends is trivial when compared to temperature (Figure S5). Even in a future low emissions scenario (SSP1-RCP2.6: average global warming across models of +1.8°C in 2048-2052; +1.9°C in 2096-2100), increases in prevalence due to historical anthropogenic climate change are projected to essentially be offset by mid-century, stabilizing around an -0.09 p.p. (95% CI: -0.41 p.p., 0.17 p.p.) projected decline across sub-Saharan Africa, relative to 2015-2020. In a high emissions scenario (SSP5–RCP8.5: +2.4°C in 2048-2052; +5.2°C in 2096-2100), we project decreases in prevalence would accelerate over time, reaching an average of -0.24 p.p. (95% CI: -0.80 p.p., 0.26 p.p.) by mid-century, and -1.90 p.p. (95% CI: -4.85 p.p., -0.07 p.p.) by the end of the century—a projected reduction that begins to approach the magnitude of some historical eradication programs.

• Download figure
• Open in new tab

Figure S3: Projected future changes in malaria prevalence driven by climate change from 2015 to 2100.

Maps show the estimated change in prevalence due to anthropogenic climate change (in percentage points) in low-emissions (SSP1-RCP2.6; A,B), moderate-emissions (SSP2-RCP4.5; C,D), and high-emissions (SSP5-RCP8.5; E,F) scenarios, projected to midcentury (2048-2052; A,C,E) or the end of the century (2096-2100; B,D,F). Projections are reported as differences relative to a present-day baseline (2015-2019). Mean estimates in panel D provide the same information as in Figure 4, but on a single color scale (i.e., no uncertainty visualization).

• Download figure
• Open in new tab

Figure S4: Historical impacts of climate change decomposed by variable.

Partial predictions of changes in malaria attributable to anthropogenic climate change are made based on all climate variables (top row), temperature (second row), flood shocks (third row), and drought shocks (fourth row). As in Figure 2, predictions based on true historical climate (blue) are compared to counterfactual predictions without anthropogenic warming (grey), relative to a 1901 to 1930 baseline. Thick lines are the median estimate across all 10,000 simulations; shading indicates the 5^th and 95^th percentiles. Plots begin in 1902 with the first full year of predictions (due to lag effects).

• Download figure
• Open in new tab

Figure S5: Future impacts of climate change decomposed by variable.

Partial predictions of changes in malaria attributable to future climate change are made based on all climate variables (top row) temperature (second row), flood shocks (third row), and drought shocks (fourth row). Projections are shown relative to the mean prevalence from 2015-2020, and as in Figure 2, line color indicates emissions scenario (blue: SSP1-RCP2.6; pink: SSP2-RCP4.5; green: SSP5-RCP8.5). Thick lines are the median estimate across all 10,000 simulations; shading indicates the 5th and 95th percentiles. Plots begin in 2016 with the first full year of predictions (due to lag effects).

Though the balance across regions will begin to shift, the geographic pattern of future changes in malaria prevalence is likely to reproduce present-day heterogeneity in impacts, as malaria transmission continues to shift along latitudinal and elevational clines in temperature (Figure 4; Figure S3). West Africa is projected to experience the most dramatic transformation, especially in a high-emissions scenario (SSP5-RCP8.5), with a projected decline of -1.04 p.p. (95% CI: -1.86 p.p., -0.36 p.p.) by mid-century, and a staggering -4.25 p.p. (95% CI: -8.78 p.p., -1.52 p.p.) decrease by 2100. Similar but shallower declines are projected in central Africa, where end-of-century reductions could reach between -0.06 p.p. (SSP1-RCP2.6; 95% CI: -0.43 p.p., 0.24 p.p.) and -1.43 p.p. (SSP5-RCP8.5; 95% CI: -4.36 p.p., 0.37 p.p.). On the other hand, localized increases in prevalence will continue in the cooler parts of the Ethiopian highlands, the greater Rift Valley region, and coastal southern Africa, potentially reaching 5 p.p. or more in some areas. The overall effect across east and southern Africa is a projected increase in prevalence, except in the highest emissions scenario (SSP5-RCP8.5), where both regions start to experience declines by mid-century, with east Africa eventually falling -0.60 p.p. (95% CI: -2.72 p.p., 0.95 p.p.) below present-day levels by 2100.

Broadly, our results suggest that the main effect of climate change mitigation will be to keep average temperatures in sub-Saharan Africa closer to the optimum range for malaria transmission. However, for many cooler localities, such as in east and southern Africa, greenhouse gas emissions reductions would prevent substantial climate change-driven increases in malaria prevalence. By mid-century, limiting global warming to below the +2°C limit in the Paris Agreement (achieved under SSP1-RCP2.6) is projected to prevent an estimated 163 cases of malaria per 100,000 children in southern Africa and 20 cases per 100,000 children in east Africa compared to an intermediate emissions scenario (SSP2-RCP4.5: +2.0°C in 2048-2052; +3.0°C in in 2096-2100). By the end of the century, these benefits would be even greater, with 505 and 33 excess cases averted per 100,000 children in southern and east Africa, respectively (Table S1). At a more local scale, these benefits could be at least an order of magnitude greater (Figure S3).

Malaria prevalence data

We use a recently published database of Plasmodium falciparum clinical prevalence in Sub-Saharan Africa⁴. 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 2015, though 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 (e.g., 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 5km 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.

For our models, we used the estimates of malaria prevalence for children aged two to ten years old, as falciparum malaria has the highest mortality in children and pregnant women. 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 aged 2-10 (Pf PR₂₋₁₀), using a catalytic conversion Muench model. For our model, we aggregated data by averaging Pf PR₂₋₁₀ at the first administrative level within-country (i.e., state or province level, or as shorthand, ADM1), using shapefiles provided by the Database of Global Administrative Areas dataset version 3.6 (www.gadm.org). This provided sufficient granularity to capture climate impacts and local heterogeneity in confounders, while ensuring sufficient data coverage within these units. This aggregation scale is supported by previous work that models this dataset at the same spatial resolution⁴.

Climate data

We used two sets of climate data in this study. The first is an observational dataset from the Climatic Research Unit (hereafter, CRU-TS; version 4.03 for model training and bias correction), which is constructed from monthly observations from extensive networks of meteorological stations from around the globe⁶⁸. CRU-TS provides landonly climatic variables at a high spatial resolution of 0.5° × 0.5° extending from 1901 to present (though our analysis is limited to the period 1901-2014). The second set of data is from ten global climate models (GCMs) selected from the sixth phase of the Coupled Model Intercomparison Project (CMIP6). 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”) 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. Shared socio-economic pathways (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 policy ^69,70. SSPs are combined with various possible future radiative forcings (representative concentration pathways; 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 (SSP1⁷¹) or a less sustainable (SSP5⁷²) economy. The third, SSP2⁷³, is a mid-way scenario, which assumes a future that mostly follows historical trends ⁷⁰. 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 apply a standard quantile-quantile (Q-Q) bias-correction^74,75 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 biascorrection, we first remap 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 perform 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 determine, 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 identify 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 after⁷⁵.

For every climate dataset (all CRU-TS and CMIP6 models), we extract the average value of monthly precipitation and temperature within each ADM1 unit. To construct polynomial variables for temperature and precipitation (see below), all data were transformed at the grid cell level prior to aggregation to the ADM1 unit; extreme precipitation cutoffs were defined at the ADM1 level and so were applied after aggregation.

Statistical model

The influence of climatic conditions on malaria prevalence has been heavily studied using transmission models based in vector ecophysiology and calibrated using laboratory experiments^6,7. The important benefit of this approach is that the mechanistic links between a particular environmental condition (e.g., temperature) and malaria prevalence in the human population, such as effects on biting rate and survival probability, can be independently isolated. However, this approach is limited in its ability to generalize to real-world contexts, where complex socioeconomic factors interact with modeled relationships based on laboratory conditions. Clinical data, which measures malaria prevalence in human populations, has been used to validate modeled results⁶, but inconsistent findings arise due to challenges in statistically isolating the role of climate from the many correlated factors influencing prevalence, such as public health interventions, drug resistance, conflict and social instability, and economic shocks ^{19,60,76,77,78}.

This study seeks to provide generalizable population-scale evidence of the malariaclimate link across sub-Saharan Africa using field-collected clinical data and a statistical approach designed to isolate changing environmental conditions from spatiotemporal confounding factors. Specifically, we draw on the climate econometrics literature ⁴⁷, which has developed causal inference approaches to quantify and project the impacts of anthropogenic climate change on a host of socioeconomic outcomes, from agricultural yields⁷⁹, to civil conflict ⁸⁰, to all-cause mortality ⁴⁸. This approach is designed to approximate controlled experiments by semi-parametrically accounting for unobservable spatial and temporal confounding factors, isolating variation in the climate system that is as good as randomly assigned ⁸¹. This approach is often referred to as “reduced-form”, as it allows for a causal interpretation of recovered relationships between socioeconomic conditions and the climate, but it does not easily enable the researcher to isolate individual mechanisms linking a changing climate to shifts in outcomes (e.g., mosquito population dynamics or parasite development rates). However, causal estimates enable counterfactual simulation in which climate is changed and all other factors are held constant; this is the exercise conducted here and in many applications of climate econometric frameworks. Moreover, these relationships can be used to calibrate more structured transmission models by providing empirical grounding from observational data.

We develop a statistical model using monthly survey-based malaria prevalence data for children aged 2 to 10 (Pf PR₂₋₁₀) covering all of sub-Saharan Africa over 115 years. Our outcome variable is the average prevalence for each first administrative unit i (e.g., province or state) in country c during month m and year t, which we denote Pf PR_icmt. We estimate prevalence as a flexible function of monthly temperature T_icmt and precipitation P_icmt variables as follows: where f (·) and g(·) represent nonlinear transformations of grid-cell level temperature and precipitation conditions, respectively, and where ℓ subscripts indicate monthly temporal lags. In our main specification, we model f (·) as a quadratic in contemporaneous average temperature, while g(·) contains a vector of dummy variables indicating whether an administrative unit’s monthly rainfall can be categorized as drought (defined as ≤ 10% of the long-run location- and month-specific mean) or flood (defined as ≥ 90% of the long-run location- and month-specific mean) during month m − ℓ. We allow for up to three monthly of lags (i.e., L = 3) for these extreme precipitation conditions in our main specification, based on hypotheses from prior literature regarding the timescales of larvae drying and of “flushing” ^50,53,82. A variety of sensitivity analyses detailed below demonstrate that key findings are robust to including lags for temperature (Figure S7), to the drought and flood cutoffs used for precipitation (Figures S8-S10), and to alternative functional forms of temperature (Figure S11).

Equation 1 uses a suite of semi-parametric spatiotemporal controls to isolate plausibly random variation in climatological conditions, following standard practices in the climate impacts literature. ^47,45 First, α_i is a vector of indicator variables for each of 853 first administrative units (i.e., “ADM1” units) across our multi-country sample. These spatial “fixed effects” control for all time-invariant characteristics of an administrative unit that may confound the relationship between temperature, rainfall, and prevalence. For example, higher altitude regions may exhibit cooler temperatures, but they also may be composed of lower-income and more geographically isolated communities with limited access to malaria prevention interventions. By controlling for mean conditions in each location, these spatial fixed effects avoid conflating climate conditions with other geographic correlates.

Second, γ_rm is a vector of region-by-month-of-year indicator variables, where regions are defined using the Global Burden of Disease (GBD) regional definitions of western, southern, central, and eastern Africa (see Figure 2 in ref.⁸³). These spatiotemporal fixed effects account for region-specific seasonality in prevalence that may spuriously relate to seasonally-varying climatological conditions. We allow these seasonal controls to vary by region because of large differences in climatological seasonality and in malaria cyclicality across sub-Saharan Africa⁸⁴, and we show below that our main findings are robust to more stringent seasonality controls defined at the country level (Figure S6). Third, h_c(·) is a nonlinear, country-specific function that controls for country-specific gradual trends that may confound the malaria-climate relationship, particularly under historical conditions of anthropogenic climate change. In our main specification, we model h_c(·) as a quadratic. Figure S6 shows that our results are robust to multiple alternative approaches to controlling for long-run trends that may vary across space.

• Download figure
• Open in new tab

Figure S6: Sensitivity of the Pf PR₂₋₁₀-temperature relationship to alternative spatiotemporal controls.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and monthly average temperature and all include fixed effects (i.e., dummy variables) at the scale of the first administrative unit (i.e., ADM1). All temperature responses are plotted relative to the model-specific temperature at which prevalence is maximized; this peak temperature is indicated in grey text and with a vertical grey line in each panel. From top-left to bottom-right, model controls are: country, year, and month fixed effects; country-specific quadratic time trends and month fixed effects; country-specific quadratic time trends and country-by-month fixed effects; country-specific quadratic time trends and intervention period and month fixed effects; country-specific quadratic time trends and intervention period and region-by-month fixed effects; country-specific quadratic time trends and intervention period and country-by-month fixed effects; region-by-year and region-by-month fixed effects; region-by-year and country-by-month fixed effects; and region-by-year and region-by-month fixed effects and country-specific linear time trends. The preferred specification used throughout the main text is the center panel, which includes country-specific quadratic time trends and intervention period and region-by-month fixed effects. In all panels, “region” refers to the Global Burden of Disease regional definitions of western, southern, central, and eastern Africa. All standard errors are clustered at the first administrative unit (e.g., province) level.

Finally, the indicator variables 𝟙{intervention 1}_mt and 𝟙{intervention 2}_mt are equal to one when an observation falls into the 1955-1969 or 2000-2015 period, respectively. These two periods saw substantial malaria intervention programs across the subcontinent, leading to considerable declines in malaria that were unrelated to changes in the climate ^4,85. These indicator variables control for shocks to prevalence during these two periods, and the coefficients δ₁ and δ₂ allow for differential effectiveness of the two distinct intervention periods. While these variables are highly statistically significant (Table S2), our main findings are robust to their exclusion (Figure S6).

Together, these set of flexible controls imply that the residual variation in temperature and precipitation events used to identify the functions f (·) and g(·) is month-to-month variation over time within the same location, after controlling for gradual country-specific trends, regional seasonality, and the aggregate effects of two substantial malaria prevention intervention programs. When reporting regression results directly (e.g., in Table S2), we cluster standard errors ε_icmt at the ADM1 level to account for serial correlation within the same location. When computing bootstrap samples (e.g., shown in Figure 2), we repeatedly re-estimate Equation 1 after block-resampling the full dataset using ADM1-level blocks to account for this same serial correlation.

Statistical model sensitivity and robustness

In this section, we describe a set of model sensitivity analyses that probe the robustness of our empirical model. Specifically, we investigate sensitivity of our key findings to: alternative spatiotemporal controls; inclusion of dynamic temperature effects; alternative definitions of extreme rainfall events; and alternative functional forms for the prevalence-temperature relationship.

Alternative spatiotemporal controls

Our preferred empirical specification in Equation 1 includes first administrative unit fixed effects (i.e., indicator variables), region-by-month-of-year fixed effects, country-specific quadratic time trends, and two indicator variables for each of two malaria intervention periods (1955-1969 and 2000-2015). Figure S6 shows that our estimated prevalence-temperature relationship is highly robust to many alternative spatial and temporal controls. All panels in this figure include ADM1 fixed effects to control for time-invariant characteristics that may confound the relationship between prevalence and temperature, but each panel varies in the additional spatial and/or temporal controls included in the regression. A tabular version of these results is shown in Table S2. While the temperature at which prevalence peaks changes slightly across model specifications, it remains within a degree of the 24.9ºC value from our preferred specification for most models, particularly those including time trends that are spatially differentiated (note that peak temperatures indicated in Figure S6 are rounded to the nearest degree for display purposes). Predictably, stringent controls, such as region-by-year and country-by-month fixed effects, tend to increase statistical uncertainty. However, overall the estimated shape and magnitude of the prevalence-temperature relationship remain robust to alternative spatial and temporal controls.

Dynamic temperature effects

Our preferred empirical specification estimates contemporaneous (within one month) and lagged (up to three months) effects of extreme rainfall on malaria prevalence, but only contemporaneous effects of temperature. While it is possible that temperature also exhibits lagged effects, we show in Figure S7 that the cumulative effect of temperature on Pf PR₂₋₁₀ is very similar whether zero, one, two, or three months of lagged temperatures are accounted for. The prevalence response to temperature does become slightly stronger with three months of lags, suggesting that our historical and future climate predictions shown throughout the main text may be somewhat conservative. However, overall these findings suggest that climate change impact predictions are unlikely to change meaningfully under different assumptions of the lag structure of temperature exposure.

• Download figure
• Open in new tab

Figure S7: Cumulative effect of contemporaneous and lagged temperature on Pf PR₂₋₁₀.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and monthly average temperature and are plotted relative to a monthly average temperature of 25^◦C. The first panel shows the effect of monthly average temperature on the same month’s average prevalence (this is the main estimate used throughout the main text). The second panel shows the cumulative effect of contemporaneous temperature and the prior month’s temperature on prevalence, while the last two columns show analogous results for two and three months of lags, respectively. In all specifications, three months of lagged precipitation extremes are included, as well as all other controls shown in Equation 1. All standard errors are clustered at the first administrative unit (e.g., province) level.

Alternative definitions of extreme rainfall events

Our main empirical specification defines drought as months for which total precipitation is less than or equal to 10% of the long-run locationand month-specific mean. Flood is analogously defined as months for which total precipitation is greater than or equal to 90% of the long-run location- and month-specific mean. Here, we investigate the sensitivity of our main findings to these definitions. To do so, we systematically vary both the drought and flood cutoff values, ranging from <1% to <20% for drought and from >85% to >95% for flood. Figure S8 shows that the relationship between malaria prevalence and temperature is insensitive to the definition of drought and flood events. Figure S9 shows that under most drought and flood definitions, extremely low precipitation events have a negative effect on prevalence with a lag of 1-2 months. However, this effect is rarely statistically significant. Figure S10 shows that extremely high rainfall events increase prevalence with a lag of 2-3 months, a result that is statistically significant and highly robust to alternative drought and flood definitions. In general, these sensitivity analyses show that our main findings are not sensitive to the specific definitions of drought and flood used in estimation of Equation 1.

• Download figure
• Open in new tab

Figure S8: Sensitivity of Pf PR₂₋₁₀-temperature relationship to alternative drought and flood definitions.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and monthly average temperature and are plotted relative to a monthly average temperature of 25^◦C. Cutoff values for drought and flood definitions are given in the titles of each panel. For example, the first panel in the upper left defines drought as monthly total precipitation that falls below 1% of the long-run location- and month-specific mean, and defines flood as monthly total precipitation that falls above 85% of the long-run location- and month-specific mean. In all specifications, three months of lagged precipitation extremes are included, as well as all other controls shown in Equation 1. The main specification used throughout the paper uses a drought cutoff of 10% and a flood cutoff of 90%. All standard errors are clustered at the first administrative unit (e.g., province) level.

• Download figure
• Open in new tab

Figure S9: Sensitivity of PfPR_2-10-drought relationship to alternative drought and flood definitions.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and contemporaneous and lagged drought events. Point estimates are given by solid circles, while vertical bars indicate 95% confidence intervals. Cutoff values for drought and flood definitions are given in the titles of each panel. For example, the first panel in the upper left defines drought as monthly total precipitation that falls below 1% of the long-run locationand month-specific mean, and defines flood as monthly total precipitation that falls above 85% of the long-run location- and month-specific mean. In all specifications, three months of lagged precipitation extremes are included, as well as all other controls shown in Equation 1. The main specification used throughout the paper uses a drought cutoff of 10% and a flood cutoff of 90%. All standard errors are clustered at the first administrative unit (e.g., province) level.

• Download figure
• Open in new tab

Figure S10: Sensitivity of PfPR_2-10-flood relationship to alternative drought and flood definitions.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and contemporaneous and lagged flood events. Point estimates are given by solid circles, while vertical bars indicate 95% confidence intervals. Cutoff values for drought and flood definitions are given in the titles of each panel. For example, the first panel in the upper left defines drought as monthly total precipitation that falls below 1% of the long-run locationand month-specific mean, and defines flood as monthly total precipitation that falls above 85% of the long-run location- and month-specific mean. In all specifications, three months of lagged precipitation extremes are included, as well as all other controls shown in Equation 1. The main specification used throughout the paper uses a drought cutoff of 10% and a flood cutoff of 90%. All standard errors are clustered at the first administrative unit (e.g., province) level.

Alternative functional forms for the prevalence-temperature relationship

Following from theoretical and laboratory-based literature (e.g., refs.^6,7), we model the prevalence-temperature relationship as quadratic. However, Figure S11 shows that this relationship is similar when more flexible functional forms are used. In particular, the temperature at which prevalence peaks changes little when higher order polynomials are estimated. Estimating higher order polynomials increases uncertainty, particularly in the tails of the temperature distribution, but point estimates are similar across the majority of the observed temperature range.

• Download figure
• Open in new tab

Figure S11: Alternative functional forms for the Pf PR₂₋₁₀-temperature relationship.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and monthly average temperature and are plotted relative to a monthly average temperature of 25^◦C. Starting in the upper left, the first panel shows the paper’s main specification, a quadratic functional form for the prevalence-temperature relationship. The second panel shows a cubic functional form, the third a quartic, and the fourth a quintic. All standard errors are clustered at the first administrative unit (e.g., province) level.

• Download figure
• Open in new tab

Figure S12: Non-linearities in the Pf PR₂₋₁₀-precipitation relationship.

All panels show the estimated relationship between malaria prevalence for children aged 2-10 and monthly cumulative precipitation and are plotted relative to a month with no rainfall. All relationships are estimated for contemporaneous monthly precipitation. All standard errors are clustered at the first administrative unit (e.g., province) level.

Predictions

In both historical and future simulations, we apply the estimated panel regression to calculate the effect of climate change on Pf PR₂₋₁₀. Our predictions capture the full range of statistical uncertainty (1,000 bootstrapped model estimates) and climate model uncertainty (10 climate models), producing a total of 10,000 estimates of historical or future impacts in any given scenario. Each of these 10,000 estimates is normalized to a long-run baseline (past: 1901-1930; present: 2015-2020) before estimates are averaged, creating an estimate of climate change impacts relative to that baseline. In our historical analyses, we only use these models to estimate changes in prevalence attributable to climate change: while the panel regression model accounts for other historical drivers through the fixed effects structure, these are not the focus of our analysis, and so we choose not to estimate total prevalence including these effects. Similarly, we elect not to make assumptions about non-climate drivers of malaria prevalence in the future, and thus do not apply the model to predict future trends in overall prevalence.

For overall trends (e.g., reported in Figures 2D, 3D, and 4D), we generated continentwide averages or four regional averages using the unweighted average of estimates for each ADM1 unit. This is a deliberate oversimplification, as we do not adjust averages based on either ADM1 units’ land area or the estimated population they contain; we made this decision based on the challenges of reconstructing historical population density at fine scales, as well as the need to otherwise make assumptions about how disease burden is allocated over space (e.g., the distribution of transmission across rural or urban areas). For similar reasons, we chose not to estimate the effect of prevalence changes on overall malaria incidence. Although some studies have attempted this using a linear conversion with total population⁸⁶, proper estimation of incidence (and the effects of treatment variables, through prevalence, on case burden) requires malaria transmission models that require substantially more demographic assumptions ⁸⁵. Future work could explore both of these methodologically-complex directions, and potentially generate finer-scale estimates of how many cases of childhood malaria, and resulting deaths, are attributable to climate change.