Projected impacts of climate change on malaria in Africa

Abstract

The implications of climate change for malaria eradication this century remain poorly resolved^1,2. Many studies focus on parasite and vector ecology in isolation, neglecting the interactions between climate, malaria control and the socioeconomic environment, including disruption from extreme weather^3,4. Here we integrate 25 years of African data on climate, malaria burden and control, socioeconomic factors, and extreme weather. Using a geotemporal model linked to an ensemble of climate projections under the Shared Socioeconomic Pathway 2-4.5 (SSP 2-4.5) scenario⁵, we estimate the future impact of climate change on malaria burden in Africa, including both ecological and disruptive effects. Our findings indicate that climate change could lead to 123 million (projection range 49.5 million to 203 million) additional malaria cases and 532,000 (195,000–912,000) additional deaths in Africa between 2024 and 2050 under current control levels. Contrary to the prevailing focus on ecological mechanisms, extreme weather events emerge as the primary driver of increased risk, accounting for 79% (50–94%) of additional cases and 93% (70–100%) of additional deaths. Most increases stem from intensification in existing endemic areas rather than range expansion, with significant regional variation in impact. These results highlight the urgent need for climate-resilient malaria control strategies and robust emergency response systems to safeguard progress towards malaria eradication.

Main

In the twenty-first century, 15 years of declining malaria burden in Africa have been followed by a decade of faltering progress. Growing optimism around new control tools is being tempered by the concurrent challenges of uncertain financing and intensifying biological threats. At this critical juncture, a question of profound importance is the extent to which climate change threatens progress against the disease and the feasibility of eradication as a mid-century goal.

The link between climate and malaria is widely accepted and extensive research has elucidated different causal pathways and sought to project future impacts^1,2. Such research has permeated local and global malaria policy formulation^6,7,8,9, but the ongoing lack of consensus on the directionality, magnitude and geographical distribution of climate change effects hinders detailed response strategies^2,10.

Climate mediates malaria vector and parasite ecology through several mechanisms. Previous studies have explored the effects of ambient temperature on Anopheles mosquito lifespan, blood meal frequency and other life history characteristics^11,12, the duration of Plasmodium extrinsic incubation period¹³, and the combined implications for vectorial capacity and malaria transmission intensity^11,12,13. Recent advances have tailored this understanding to the most important malaria vectors in Africa—members of the Anopheles gambiae complex¹⁴—and the invasive vector Anopheles stephensi¹⁵. Other work has focused on understanding the way climate affects the abundance of adult mosquitoes by influencing larval habitat and its suitability to sustain development of the larval stages^16,17.

Climate change and malaria in context

As understanding of climate effects on malaria transmission ecology has developed, many studies have extrapolated those mechanisms under future climate scenarios to predict how malaria may respond to climate change³. Some studies have incorporated temperature effects only^18,19, whereas others have used more comprehensive empirical or mechanistic models to capture temperature, rainfall and humidity effects on both larval and adult mosquito stages^4,20,21. Most commonly, these efforts have aimed to predict changes in the geographical (or seasonal) boundaries of transmission^21,22,23. However, many areas of the world are already suitable for malaria transmission yet have none, whereas within existing boundaries transmission can vary by several orders of magnitude. Understanding boundary changes therefore provides only partial insight into climate change impacts on malaria.

Other studies have sought to use mechanistic models to predict climate change impacts on transmission in absolute terms, on the basis of either extrapolated parametrizations from laboratory or field observations^20,21, or calibration to local observations^24,25. However, such models show limited ability to reproduce historical endemicity²⁶ or observed patterns of malaria infection prevalence across Africa, casting doubt on their reliability for projecting future changes in those patterns^17,20,22.

Nearly all existing projections share a central limitation: although they explore climate effects in isolation, they do not adequately account for non-climate determinants of malaria trends. Rising drug resistance in the 1990s contributed to increasing malaria trends in many locations, including those studied intensively for climate change effects^23,27. Subsequently, aggressive scale up of effective interventions between 2000 and 2015 almost halved mean infection prevalence across the continent, albeit unevenly²⁸. This, along with rapid urbanization and socioeconomic development, has reshaped the modern-day landscape of malaria risk, driven long-term changes independently of climate trends and mean that identical climate conditions can host very different levels of malaria transmission^29,30. Failing to account for these factors prevents an accurate assessment of the role of climate—and, by extension, climate change—in shaping future malaria risk.

The built environment and malaria control efforts not only mediate transmission independently of climate but also provide additional pathways by which weather and climate can impact malaria. Extreme weather events such as floods and cyclones damage homes and infrastructure, disrupting access to healthcare, protective housing and malaria control. Malaria surges linked to recent extreme weather events in Africa and Asia have been documented widely^31,32, including causal studies on the impact of disrupted malaria control, even after substantial emergency interventions³¹. In Africa, climate change is predicted to lead to more frequent and severe floods, and more severe southern Indian Ocean cyclones³³. Although some attention has been given to vector-borne disease and extreme weather in localized studies³⁴, previous climate–malaria work has focused almost exclusively on ecological mechanisms. The disruptive effects that intensifying extreme weather events might have on malaria, and vector-borne diseases more broadly³, and the control of such diseases across Africa has not previously been assessed quantitatively.

Projecting climate impacts on malaria

Here we provide projections of how both ecological and disruptive climate-change effects might affect malaria in Africa. Crucially, we first characterize the climate–malaria relationship in the context of the other key determinants of malaria risk, based on 25 years of comprehensive data on malaria infection prevalence, intervention coverage and socioeconomic conditions. A schematic overview of our analytical framework is shown in Extended Data Fig. 1. Bias-corrected and downscaled Coupled Model Intercomparison Project Phase 6 (CMIP6) member global climate model (GCM) outputs obtained from the NASA Earth Exchange Global Daily Downscaled Projections provided future projections of climatic variables to 2050 (ref. ³⁵), and were calibrated to observed climate data^36,37,38 from the co-observed 2000–2014 period, generating consistent time-series of climatic variables from 2000–2050 at monthly time-steps and 5 × 5 km spatial resolution. Between-GCM uncertainty and variability were accounted for using an ensemble of CMIP6 members—14 models for evaluating ecological impacts, and a tractable subset of three for disruptive impacts. We focused analysis on the ‘middle of the road’ scenario SSP 2-4.5, designed to be broadly consistent with current international pledges on reduced greenhouse gas emissions⁵.

Ecologically driven malaria impacts were assessed by first using mechanistic models to transform the processed climate data into two indices representing the climate–malaria interface: (1) the effects of temperature on mosquito and parasite lifecycles and their interaction and (2) the interaction of rainfall, humidity and temperature to determine the relative availability of transient larval habitat for mosquito oviposition and larval development. Use of GCM outputs with monthly resolved projections enabled appropriate propagation of changing regimes of seasonal, inter-annual and inter-decadal climate variation. Unlike previous seasonal models¹⁸ focusing only on temperature effects, this approach allows seasonality to be characterized by the intersection of temporally varying temperature, precipitation and humidity conditions. Second, for the period of observational record in the training data (2000–2022), the two climatic indices were used as geotemporal predictor variables, along with gridded data on mosquito relative species abundance³⁹, permanent larval habitat availability, geotemporal estimates of vector control coverage⁴⁰, seasonal malaria chemoprevention (SMC), antimalarial drug treatment⁴¹ and improved housing⁴² in a hierarchical Bayesian geotemporal model fitted to 49,994 geo-located observations of malaria infection prevalence (Plasmodium falciparum parasite rate (PfPR): age- and diagnostic-standardized parasitaemia rates in 2- to 10-year-old children, henceforth PfPR_2–10) collated across Africa by the Malaria Atlas Project⁴¹. This framework extended an established approach²⁸ and estimated empirically the relative and absolute effects of each predictor variable, including their interactions, on malaria transmission. Third, holding malaria control coverage and socioeconomic metrics at present-day levels, the fitted model was used to generate both SSP 2-4.5 and counterfactual (that is, projecting present-day conditions with no future climate change) PfPR_2–10 scenarios for each climate ensemble member from 2024 to 2050, which were then differenced to generate estimates of climate change impact (see detailed scenario definitions in Extended Data Table 1).

Disruption-driven malaria impacts were assessed by first simulating plausible scenarios of flooding and cyclone events in sub-Saharan Africa consistent with both past observations^43,44 and GCM-projected future climate using statistical models calibrated and validated to historical events. For fluvial and pluvial flooding, the frequency, duration and extent of observed flood events were associated with a range of climate variables to generate a set of geotemporal catalogues of simulated future flood events under both counterfactual (without future climate change) and SSP 2-4.5 conditions to 2050. Similarly, the statistical relationship between climate variables and past Indian Ocean cyclone intensity and track morphology was characterized using methodology and datasets from a recently updated storms model⁴⁵ that was also then projected to generate SSP 2-4.5 and counterfactual event catalogues to 2050.

Although evidence exists on the link between extreme weather events and vector-borne disease^46,47, including malaria, impacts have almost never been measured directly. We assembled both qualitative and quantitative records of past extreme weather events in malaria-endemic settings that detailed the type and magnitude of damage and disruption experienced, and supplemented these with insights from 34 stakeholder interviews, capturing first-hand accounts of malaria impacts on the ground, including from representatives of humanitarian response agencies, national malaria programmes and global health agencies. This mixed-methods approach identified four primary pathways by which extreme weather events can impact malaria: by damage or disruption to the protective effects of (1) improved housing, (2) indoor vector control tools or (3) access to antimalarial treatment, or from changes to Anopheles larval habitat availability due to flood water (Extended Data Table 2). For (1)–(3), we used the compiled evidence (Extended Data Table 3) to define plausible levels of disruption by event type, duration and intensity—distinguishing between acute disruption in the immediate aftermath, and a period of persisting disruption before return to pre-event conditions (see Extended Data Fig. 2 for a schematic overview of our approach). To reflect the inherent variability and uncertainty in these disruption parameters, we also generated ranges spanning 50% to 150% of the central consensus values. A broad uncertainty range was considered appropriate because the scarcity of observational data precluded a more formal quantification of variation in these disruption parameters. These uncertainty ranges were propagated into all downstream modelling steps and the final results. By combining the disruption parameters with the simulated extreme events, we generated modified versions of the geotemporal data layers for vector control coverage, antimalarial drug treatment and improved housing that incorporated disruptions under both SSP 2-4.5 and counterfactual scenarios to 2050. Larval habitat impacts were captured by the habitat suitability model described earlier. The fitted hierarchical Bayesian geotemporal model was then used to reproject PfPR_2–10 with these disruptive impacts, both separately and in combination with the ecological impacts, which were then differenced from counterfactual scenarios without future climate change (Extended Data Table 1). Uncertainty in the damage parameter ranges was combined with GCM model ensemble spread to derive overall projection ranges (Extended Data Fig. 6). All projections of PfPR_2–10 were converted into estimates of malaria clinical incidence using an established natural history model⁴⁸ and with gridded future population projections consistent with SSP 2-4.5 (ref. ⁴⁹). Mortality projections were derived from projected untreated cases and World Health Organization estimates of untreated case-fatality rates⁵⁰.

Ecological versus disruptive impacts

We project that, considered in isolation, the ecologically driven impacts of climate change on malaria transmission would lead to minimal overall change in Africa by 2050 under SSP 2-4.5, with a continent-wide population-weighted mean percentage point increase in PfPR_2–10 due to ecological impacts of just 0.12% relative to the counterfactual without future climate change (14-model ensemble spread, −0.03% to 0.33%). However, this continental aggregation masks extensive geographical variation (Fig. 1a–c), with much larger local and regional impacts predicted. We project that warming under SSP 2-4.5 would increase malaria risk in regions where lower temperatures currently constrain transmission: the belt of lower latitude Southern Africa, including Angola, Zambia and southern Democratic Republic of Congo (DRC), as well as highland areas in Ethiopia, Kenya, Rwanda, Burundi and eastern DRC (Fig. 1c). Conversely, warming would drive reduced transmission in Sahelian regions as temperature regimes would increasingly exceed optimal ranges for Anopheles survival (Fig. 1b). Changing temperature suitability (Fig. 1d) rather than larval habitat availability (Fig. 1e) was identified as the dominant ecological mechanism, although this was driven partly by the greater GCM ensemble divergence for rainfall projections versus those for temperature (Extended Data Fig. 3).

Fig. 1: Projected ecologically driven impacts of climate change on malaria transmission under SSP 2-4.5.

a, Mapped percentage point change in PfPR_2–10 due to ecologically driven impacts of climate change. Values compare annual median across 2040–2049 under SSP 2-4.5 scenario (ensemble mean of 14 downscaled and bias-corrected CMIP6 members) to counterfactual (no future climate change) scenario, with intervention coverage and socioeconomics held constant at present-day levels in both scenarios. Large water bodies and out-of-scope countries are masked in dark grey. b,c, Time-series of population-weighted 3-year rolling mean change in PfPR_2–10 from 2019–2022 baseline in the Sahel (b) and areas of mean elevation above 1,500 m (c), with ribbons representing ensemble 25th–75th (darker green), and 10th–90th percentile (lighter green) and inset maps showing geographical subsets. d,e, Decomposition of ecological impacts due to changing temperature (d) and changing larval habitat (e) suitability, with colour ramping identical to a. Individual ensemble member summaries are shown in Extended Data Fig. 3. Administrative boundaries were obtained from the Malaria Atlas Project (https://data.malariaatlas.org/)⁴¹, under a CC BY 3.0 licence.

Our simulated flood event catalogues reflected a net increase in flooding across Africa due to climate change, with area-days flooded 13% (ensemble spread 9–17%) greater in the 2040s under SSP 2-4.5 versus present-day conditions (Extended Data Fig. 4). We projected Indian Ocean cyclones to shift in severity under SSP 2-4.5 versus present-day conditions, with fewer category 1–4 events but more frequent category 5 (Extended Data Fig. 4). Although there is substantial uncertainty reflected in the ensemble spread, our projections are consistent with current consensus and reflect expectations of increased atmospheric moisture transports and intensification of the water cycle^9,33. We project that, under SSP 2-4.5, the disruptive impact of intensifying extreme weather events in the 2040s could, without mitigation measures, lead to PfPR_2–10 increasing across 24% of the land area of malaria-endemic Africa, associated mainly with main river systems and the cyclone-prone coastal regions of southeast Africa (Fig. 2a).

Fig. 2: Projected disruptive and combined disruptive-plus-ecological impacts of climate change on malaria transmission in the 2040s.

a, Mapped percentage point change in PfPR_2–10 due to disruptive impacts of climate change only. Values compare PfPR_2–10 median across 2040–2049 impacted by simulated extreme weather events consistent with the SSP 2-4.5 scenario (ensemble mean of three downscaled and bias-corrected CMIP6 members) versus a counterfactual scenario with extreme weather events projected at present-day levels. Intervention coverage and socioeconomics otherwise held constant at present-day levels. Large water bodies and out-of-scope countries masked in grey. b, Mapped percentage point change in PfPR_2–10 due to combined disruptive and ecological impacts of climate change. Values compare PfPR_2–10 median across 2040–2049 under the SSP 2-4.5 scenario versus the counterfactual (no future climate change) scenario. Administrative boundaries were obtained from the Malaria Atlas Project (https://data.malariaatlas.org)⁴¹, under a CC BY 3.0 licence.

The combined ecological and disruptive impacts of climate change on PfPR_2–10 projected to the 2040s are shown in Fig. 2b and Extended Data Fig. 5. In several settings—notably Uganda, eastern DRC, Zambia and Angola—the ecological and disruptive effects compound, with ecologically driven increases in underlying transmission exacerbated by more frequent extreme weather events that further expose those vulnerable populations. Across floodplains in the Sahel, meanwhile, projected decreases in ecological suitability are negated by more frequent and severe flooding, for example along the Niger River in Mali and Nile in South Sudan. In combination, we project that ecological and disruptive impacts in the 2040s under SSP 2-4.5 would lead to increased malaria transmission for 67% of Africa’s population, with around 73 million exposed to a percentage point increase in PfPR_2–10 in excess of 2% (Extended Data Fig. 5).

Increased malaria due to climate change

Over the next 25 years, we project that climate change—through combined ecological and disruptive impacts—could lead to 123 million additional clinical cases of malaria in Africa (projection range 49.5–203 million) under SSP 2-4.5 (Fig. 3a). By the 2040s, this equates to an increase of 0.4–2.6% in case incidence rate, relative to the counterfactual without future climate change. The growing disruptive impact of more frequent and severe extreme weather events contributes 79% (projection range 50–94%) of this projected rise in case incidence—more than three times greater than the contribution of ecologically driven impacts (21%, 6–50%). Of the disruption-driven impacts to case incidence, those resulting from reduced access to malaria treatment form the largest component (37.8% of total, ensemble spread 34.2–41.3%), followed by damage to housing (23.4%, ensemble spread 21.3–27.1%) and disruptions to vector control (14.9%, ensemble spread 14.6–15.5%) (Fig. 3a). Figure 3b maps the cumulative change in clinical cases by 2049 for each 5 × 5 km grid cell. Some locations with large projected rises in transmission rates occur in relatively sparsely populated regions—for example, central Angola and western Zambia—and hence incur only modest increases in cases. Elsewhere, substantial climate impact coincides with regions of dense populations at risk, most notably southern and central Nigeria, and the African Great Lakes region encompassing parts of Kenya, Uganda, eastern DRC, Rwanda and Burundi. Assuming present-day untreated case fatality ratios⁵⁰, our projected changes to case incidence and treatment rates would equate to 532,000 (195,000–912,000) additional malaria deaths over the same period—a relative increase in annual mortality rates of 0.9–5.4% compared with the counterfactual in the 2040s.

Fig. 3: Projected cumulative impact on clinical malaria incidence 2024–2049 from climate change under SSP 2-4.5.

a, Projected continent-wide cumulative impact on cases by year, with decomposition showing distinct contributions of the different ecological and disruptive drivers of climate change impact. Grey bars denote projection range, where this is calculated as the 10th and 90th percentile of the model ensemble across all GCMs and disruption parameter ranges. b, Mapped cumulative impact on cases with colour scale reflecting degree of GCM ensemble consensus. Large water bodies, national parks with extremely low or zero population, and out-of-scope countries are masked in dark grey. All cumulative impacts are relative to 2022 case counts, aligned to national totals reported in the 2023 World Malaria Report⁵⁰. Administrative boundaries were obtained from the Malaria Atlas Project (https://data.malariaatlas.org)⁴¹, under a CC BY 3.0 licence.

Many preceding studies have focused on identifying climate change-driven changes to the boundaries of transmission suitability. However, we project that the great majority of additional cases would occur in settings already suitable for transmission, with only 0.05% (33,000 cases, three-GCM ensemble spread 20,000–51,000, 0.03–0.07%) of additional cases in the 2040s occurring in places currently outside suitability limits.

Discussion

This analysis examined the potential threats posed by both ecological and disruptive climate-change impacts on malaria transmission and burden in Africa and quantified the plausible magnitude of these impacts in absolute terms while placing climate mechanisms in the context of other key factors that mediate malaria risk. Although the projected impacts on transmission are ostensibly modest, we have shown that, when translated into impact on clinical incidence and deaths over the coming decades, the implications of climate change become substantial, potentially resulting in hundreds of millions of additional cases and hundreds of thousands of additional deaths over the next 25 years. In the wider context of plans for accelerated burden reduction and eradication over that timescale, such climate impacts would be consequential and demand multi-sectoral policy responses.

In our analysis of ecologically driven climate impacts, we have taken current state-of-the-art understanding of the main biophysical mechanisms involved and, uniquely, used detailed historical malaria, intervention and broader contextual data to estimate the absolute contribution of these mechanisms to realized malaria transmission. We have also demonstrated that a narrow focus on changing transmission boundaries ignores around 99% of likely climate impact in Africa, which occurs in regions already experiencing transmission.

Unlike the analysis of ecological impacts, our modelling of disruption-driven impacts had very limited precedent in the literature to build upon or compare against and the paucity of data necessitated a more heuristic approach to quantifying likely impacts. Nonetheless, by compiling quantitative and qualitative evidence, including the first-hand accounts of those witnessing and responding to extreme weather events in Africa, we have identified mechanisms and plausible magnitudes of impact, with appropriate representation of uncertainty, that probably reflect real-world experience. By combining these effects with simulated future events consistent with GCM projections, we provide a systematic exploration of the impacts of worsening, climate-change-associated extreme weather events on malaria. We identify these disruptive impacts to be around three times larger than ecologically driven effects.

This analysis should be interpreted in the context of its design, assumptions and limitations. (1) Our results reflect multi-model ensemble means across decadal timespans. This choice reflects a balance between reducing artefactual inter-model and inter-annual variation while addressing the medium-term timescales relevant to urgent malaria policy needs. Decadal summaries are, by definition, also subject to inter-decadal climate variability known to influence malaria⁵¹, but an ensemble approach reduces their artefactual impacts. (2) Some sources of uncertainty are not captured, including the following: uncertainty in the future climate projections (which is accounted for only partially through ensembling); non-climate and other unaccounted-for drivers of future malaria trends; uncertainties in future population projections; and imprecision in our projected flood and cyclone extents, severities and durations. The projection ranges are not, therefore, intended as formal statistical intervals, providing indicative measures of uncertainty rather than probabilistic statements. (3) We limited our analysis of disruptive impacts to floods and cyclones, for which the evidence base of impact on transmission is best established^31,32. Future climate change will conceivably impact malaria through a wide range of indirect mechanisms, such as food insecurity, loss of livelihoods, conflict, health system stressors and economic disruption. Many of these drivers also imply increases in climate-driven migration, which may have profound implications for malaria burden as populations potentially become exposed to different transmission risks and reduced access to prevention and treatment infrastructure. Although such mechanisms have not been included in this analysis, the importance of considering broader, cross-sectoral climate impacts on malaria and other vector-borne diseases has been emphasized³ and concerted future research will be required in this domain. (4) We deliberately hold constant present-day levels of transport and healthcare infrastructure, housing quality and malaria control—accounting for direct climate impacts on these determinants but not background trends. Thus, our projections allow exploration of climate change effects but are not intended as forecasts of future conditions. In reality, socioeconomic trends across future decades will inevitably exert substantial and uneven impact on malaria—both directly and by mediating climate–malaria relationships. (5) Although the stacked modelling approach used here allows flexible response–predictor relationships and demonstrates good out-of-sample predictive performance, there is currently limited precedent for fully characterizing uncertainty in stacked models, particularly when applied to spatially correlated data. Further theoretical work is needed to establish rigorous uncertainty quantification in stacked and other ensemble spatial modelling frameworks.

The framework we present here addresses several of the recognized gaps³ in preceding climate–malaria analyses, including the combined use of mechanistic and correlative approaches, the consideration of extreme weather events and the propagation of ensembled climate models. Taken as a whole, our findings demonstrate the potential of climate change to substantially hinder malaria reduction and eradication over the coming 25 years, while emphasizing that the ultimate outcome will hinge on the effectiveness of malaria control and the resilience of the health systems delivering it. Although the projected scale of impacts from gradually changing transmission ecology could probably be counteracted with modest additional intervention, increasing damage and disruption to control efforts from extreme weather poses a more profound threat. Mitigating this threat will require renewed focus on climate-resilient control strategies at international and local levels. This might include, for example, investment in more climate-resilient health and supply chain infrastructure, enhancement in emergency early warning and response mechanisms, decentralization of vertical and horizontal health service delivery to foster local resilience, as well as locally tailored use of new tools less vulnerable to climate disruption⁵².

Eradicating malaria in the first half of this century would be one of the greatest accomplishments in human history. Accelerating climate change poses numerous threats to this ambition. In this context, climate-proofed eradication strategies will demand ongoing vigilance, proactive planning, engaged communities and sustained financing, all within the broader framework of robust health systems and climate-resilient societies.

Methods

Our analysis framework comprised nine main stages, summarized in Extended Data Fig. 1.

Preparation of consistent geotemporal climatologies, 2000–2050

Historical climate data

Climate data were obtained from the Climate Hazards Centre³⁶ and Climate Research Unit gridded Time Series³⁷, downscaled by Worldclim³⁸. Data were gap-filled⁵³, aligned to a standard 5 × 5 km reference grid and aggregated to monthly time-steps. Details of all historical climate data used in this study are provided in Supplementary Information Table 1.

CMIP6 projections

Projections of future climate under SSP 2-4.5 were obtained from the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6)³⁵. These data consist of downscaled and bias-corrected daily CMIP6 multi-model ensemble outputs for historical (2000–2014) and future projection (2015–2050) eras. Data were aggregated to monthly resolution before applying the delta method⁵⁴ to provide a final calibration to historical climate data described above. To account for between-model uncertainty, 14 models were processed and used for projection of ecologically driven climate impacts (Supplementary Table 2), whereas a thinned subset of three models (ACCESS-CM2, EC-Earth3-Veg-LR and MPI-ESM1-2-LR) was used for the additional analysis of disruptive climate impacts.

Modelling of historical and projected climate suitability indices

The assembled climate data were used in two mathematical models to develop two geotemporal suitability indices: (1) a temperature suitability index (TSI) tracking relative vectorial capacity and (2) a larval habitat suitability index (HSI) measuring relative availability and potential productivity of mosquito larval breeding sites.

Temperature provides both an upper and lower constraint on malaria transmission, reflecting the ectothermic nature of mosquito and parasite lifecycles. Mathematical models linking temperature to relative vectorial capacity are well established and described elsewhere⁵⁵. Here we updated a degree-day-based framework^56,57 to include recently published data on the A. gambiae complex¹⁴. The resulting temperature suitability curve is nonlinear, peaking at 26.4 °C.

To describe the relationship between local rainfall, temperature and humidity and the resulting availability of habitat for oviposition and larval development we discretized a Clausius–Clapeyron-based model of habitat availability used in an established mechanistic model of malaria⁵⁸ as follows: let R_t denote rainfall volume at time t, so that transient larval habitat is given by the recursion:

$${\rm{Habitat}}(t)={R}_{t}+\frac{1}{{\delta }_{t}}{\rm{Habitat}}(t-1),$$

where δ_t is the time-dependent temperature and humidity-dependent evaporation rate, which is a function of temperature T_t, humidity H_t and a physical constant C:

$${\delta }_{t}=C\exp \left\{-\frac{1}{{T}_{t}+273.15}\right\}\sqrt{\frac{1}{{T}_{t}+273.15}}(1-{H}_{t}).$$

Then the expected duration of habitat at time t is:

$${\mathrm{dur}}_{t}=\frac{1}{{\log }_{2}}\frac{1}{{\delta }_{t}},$$

and we may approximate Habitat(t) as HSI(m) for a month m of length |m|:

$$\begin{array}{l}{\rm{HSI}}(m)=\sum _{d\in m}\frac{{R}_{d}}{|m|}\,\min ({{\rm{dur}}}_{m},|m|-d)\\ \,\,\,+\sum _{d\in m-1}\frac{{R}_{d}}{|m-1|}\,\max (0,{{\rm{dur}}}_{m}-|m|+d){\mathbb{I}}({{\rm{dur}}}_{m-1} > |m-1|-d)\end{array}$$

where ${\mathbb{I}}$ denotes an indicator variable equal to 1 when the condition is true and 0 otherwise. We augmented this transient larval habitat suitability with a ‘permanent larval habitat’ layer derived from Tasselled Cap Wetness^53,59 observations adjusted for local rainfall.

To visually examine the empirical signal of these two climatic suitability indices, we binned PfPR observations into deciles by TSI and HSI (Extended Data Fig. 7). Subsetting observations to those in rural settings, and then further to areas of low insecticidal bednet (ITN) coverage, the variation in PfPR associated with changing TSI and HIS becomes more pronounced.

Preparation of historical geotemporal housing, intervention and contextual data

Malaria control

The scale up of malaria control is responsible for most of the decline in burden in sub-Saharan Africa since 2000 (ref. ²⁸). Geotemporal estimates of historic ITN, indoor residual spray (IRS), SMC and antimalarial treatment coverage were obtained from the Malaria Atlas Project. To account for the nonlinearity in the ITN–PfPR_2–10 relationship, ITN coverage was transformed using a pre-defined parametric interaction with estimated pre-intervention-era parasite rate²⁸. The resulting functional form of the ITN–PfPR_2–10 relationship was consistent with biological understanding as encoded in mechanistic malaria models^60,61. This study did not seek to model possible future trends in intervention coverage. Instead we defined a baseline intervention coverage reflecting present-day levels, and this was used for all future scenarios (with or without disruption due to extreme weather events). To account for the campaign-based nature of ITN distribution we took a 4-year average across 2019–2022 as baseline. For treatment coverage, delivered horizontally, 2022 was taken as baseline for future projections.

Relative abundance of vector species

Malaria ecology varies by mosquito species, with the relationship between the environment and transmission expected to vary between settings with different dominant vectors, even after controlling for interventions and socioeconomic factors. To account for this, we fit species-specific terms in the model, with predictions based on relative abundance of the three dominant vectors in sub-Saharan Africa: A. gambiae (s.s. and coluzzi), Anopheles funestus and Anopheles arabiensis. Estimates of relative abundance of each species were obtained from ref. ³⁹, providing for each 5 × 5 km grid cell a three-way weighting of malaria vector species. We omit explicit modelling of the invasive Anopheles stephensi vector but acknowledge this mosquito represents a substantial emerging threat.

Socioeconomics

Improved housing has been shown to reduce risk of malaria infection⁶² and is a key mechanism by which socioeconomic development impacts malaria independently of direct malaria investment. We updated existing estimates of the prevalence of improved housing⁴² using data on characteristics of 1,083,386 households across Africa⁶³ and predictors that include gridded data on population density⁴⁹, gross domestic product⁶⁴, land cover⁶⁵ and travel time⁶⁶. To account for the lifespan of buildings, a monotonicity constraint (increasing) was applied. We did not seek to model future trends in improved housing, and 2022 was used as the present-day baseline for projections.

Preparation of historical malaria response data

A total of 49,994 geo-located observations of PfPR were collated from Demographic and Health Surveys (DHS) Program and other nationally representative surveys⁶³ and systematic literature review⁴¹, representing 2.54 million people tested between 1995 and 2021 across 41 African countries. These data were standardized for age (to 2–10 years) and diagnostic type (to light microscopy), consistent with established approaches to modelling these data⁶⁷.

Characterizing historical climate, housing, intervention and other contextual effects on malaria

We used stacked generalization⁶⁸ to regress the predictor variables onto PfPR observations. This method, which ensembles out-of-sample predictions from several ‘level 0’ models as predictors in a final geostatistical generalizer, has been shown to outperform standard model-based geostatistics when applied to PfPR data⁶⁸.

Three ‘level 0’ models were used in this analysis:

• (1) Linear model: for observed PfPR y,

  $$\begin{array}{c}\mathrm{elogit}\,{y}_{s,t} \sim \mathrm{Normal}({{\rm{\mu }}}_{s,t},{\sigma }_{{\epsilon }}^{2})\\ {{\rm{\mu }}}_{s,t}={U}_{c[s]}+\tilde{\beta {\prime} }{X}_{s,t}^{\mathrm{climate}}+\tilde{\gamma {\prime} }{X}_{s,t}^{\mathrm{contextual}}\end{array}$$

  where U_c[s] was a country-specific intercept, ${X}_{s,t}^{{\rm{climate}}}$ was a matrix with columns (i) standardized as inh-transformed HSI, cumulative across 2- and 3-month lags; (ii) standardized TSI at 1-month lag, (iii) their multiplicative interaction; (iv) and (v) A. arabiensis relative abundance weighted versions of (i) and (ii); and finally (vi), (vii) A. funestus relative abundance weighted versions of (i) and (ii) (A. gambiae (s.s. and coluzzi) was taken as our reference species for terms (i) and (ii)). ${X}_{s,{t}}^{{\rm{contextual}}}$ was a five-column matrix consisting of prevalence of improved housing, the ITN interaction term, IRS coverage, access to effective treatment with an antimalarial and SMC coverage. ${\sigma }_{{\epsilon }}^{2}$ denotes variance of the Gaussian noise term. ${\rm{elogit}}$ denotes the empirical logit.

• (2) Generalized additive model: model formula as for meta-model (1), with A. funestus as reference species, and the linear terms for A. gambiae relative abundance weighted HSI, housing, IRS and SMC replaced with penalized cubic splines.

• (3) Generalized boosted regression: model formula as for meta-model (1), with interaction term removed and replaced with species-specific terms. Monotonicity constraints were placed on the model terms to prevent spurious results. A total of 1,000 trees were fit, with an interaction depth of four.

These three models were fit to the PfPR_2–10 dataset described above, avoiding overfitting by using fivefold cross-validation to generate out-of-sample predictions to be used as fixed effects when training the level 1 generalizer model. For tractability, we modified the stacked generalization approach by, for each level 0 model, generating these hold-out predictions with interventions set to zero. That is, each of the level 0 models provided an (out-of-sample) predictor representing estimated risk in the absence of interventions.

Each intervention class was then included in the geostatistical primary model as a fixed effect. For location s and time t the generalizer model can be expressed, for PfPR y_s,t, as:

$$\begin{array}{c}\mathrm{elogit}\,{y}_{s,t} \sim \mathrm{Normal}({\eta }_{s,t},{\sigma }_{{\epsilon }}^{2})\\ {\eta }_{s,t}=\mathop{\sum }\limits_{i=1}^{3}{\beta }_{i}^{{\prime} }{X}_{i,s,t}+\mathop{\sum }\limits_{j=1}^{4}{\gamma }_{j}^{{\prime} }{I}_{j,s,t}+{Z}_{s,t}\\ {Z}_{s,t} \sim {\mathcal{G}}{\mathcal{P}}\,(0,{\Sigma }_{\mathrm{space}}\otimes {\Sigma }_{\mathrm{time}})\\ {\Sigma }_{\mathrm{space}} \sim \mathrm{Mat}{\rm{ \acute{{\rm{e}}} }}\mathrm{rn}(5/2),\mathrm{PC}\,\mathrm{prior}\,:P\left(\mathrm{range} < \frac{7{\rm{\pi }}}{180}\right)=0.1,\\ P({{\sigma }}_{\mathrm{space}} > \,\log (1.1))=0.01\\ {\Sigma }_{\mathrm{time}} \sim \mathrm{AR}(1)(\rho ),\mathrm{PC}\,\mathrm{prior}\,:P(\rho > 0.7)=0.99,\\ P({{\sigma }}_{\mathrm{time}} > 0.05)=0.01\\ {\beta }_{i}^{{\prime} } \sim N(0,1)\\ {\gamma }_{1}^{{\prime} } \sim N(4,0.1)\\ {\gamma }_{2}^{{\prime} } \sim N(0.5,0.1)\\ {\gamma }_{j}^{{\prime} } \sim N(0.5,1),j=3,4\end{array}$$

where X_s,t are the out-of-sample zero-intervention predictions forming the level 0 stack, I_s,t is the vector of malaria control coverage (interacted with pre-intervention-era PfPR, in the case of ITNs) at pixel s at time t, Z is spatio-temporal random field with separable Matérn-5/2 spatial and AR(1) temporal kernel. The meta-learner slopes β′ were non-negative—a sufficient condition for stacked generalization⁶⁸. Following previous approaches to modelling PfPR, we used a Gaussian likelihood and empirical logit transform to ensure well-specified posterior distributions⁶⁹. The geostatistical model was fit in R 4.2.0 using INLA v.22.12.16 (refs. ^70,71). Fitted parameters for both the level 0 models and geostatistical generalizer model are given in the Supplementary Information.

This modelling approach estimates the empirical associations between each predictor and PfPR_2–10. Counterfactual predictions may then be constructed by modifying covariates of interest, holding other variables constant, and using the fitted parameters to generate predictions of PfPR_2–10, which can then be differenced from a baseline prediction in which all variables were held constant. By repeating the procedure for each predictor and combination of predictors, we estimate the change in response attributed to change in covariates, conditioned on the model structure and observed data.

Validation

In-sample observed versus fitted correlation was 0.86 and mean absolute error was 8.5% at cluster and 1.1% at DHS survey aggregate. Verification of out-of-sample predictive ability using fivefold cross-validation yielded out-of-sample correlation of 0.83, mean absolute error of 9.5% (2.5% DHS survey out-of-sample aggregate). Further validation details are provided in Supplementary Information section 5.2.

Generation of future scenarios of extreme weather events

Scenario-based projections of flood events

Random Forest models were developed to predict flood occurrence, extent and duration. The training set consisted of historic flood events extracted at 230-m spatial resolution from Floodbase⁴³, aggregated to Pfafstetter level 4 basin and over-sampled to reduce bias arising from data imbalance. A binary classifier was trained on historic level 4 basin flood occurrence using predictors that included rainfall and Atlantic and Indian Oceans sea-surface temperatures (see Supplementary Information Table 7 for a full description of 27 predictors).

Flood extent (in square kilometres) and duration were modelled probabilistically with similar predictors (see Supplementary Tables 8 and 9 for full description).

The performance of the three flood models was assessed using a 30% test set. The frequency model correctly classified 82% of occurrences and 85% of non-occurrences (area under the curve 0.84), whereas the duration and extent models had R² values of 0.83 and 0.81, respectively.

We calculated GCM-specific future projections of flood occurrence, duration and extent by level 4 basin for 2024–2049. For each GCM, the simulated period 2024–2026 was resampled to generate future counterfactual scenarios representing present-day climate flood frequency and extents. To downscale the predicted level 4 basin-level extents, high-resolution occurrence data of historic floods⁴³ were combined with a hydrological model⁷² to obtain, for each grid cell, flood propensity scores. For each flood event, grid-cells were then flooded from highest to lowest flood propensity until the predicted extent was reached. Projections were ceteris paribus, with land use variables held constant.

Scenario-based projections of cyclones

Historical Indian Ocean cyclone data (track morphology, start and end date, wind speeds) were obtained from IBTrACS⁴⁴. Storms in the IBTrACS database coming within 50 km of the coast of Africa were included, yielding a training set of 192 tropical depressions, storms and cyclones since 1980. Future scenarios of cyclone genesis and trajectories were generated using the Imperial College Storm Model (IRIS) dataset⁴⁵—a 10,000-year synthetic dataset of statistical characteristics of cyclones, adjusted with climate data from the downscaled and bias-corrected CMIP6 models. Following the IRIS method, we modelled cyclone generation events (commencing from the point of lifetime maximum intensity (LMI)), with probability of occurrence at each location in the southern Indian Ocean modelled as Poisson distributed with spatial variation learned using coordinates of historic LMI locations. From these LMIs, synthetic tracks were generated by perturbing historical tracks using forecast cone uncertainty.

The maximum sustained wind speeds at LMI were calculated using climate data and thermodynamic constraints, including potential intensity. After LMI, the model simulated intensity decay separately for ocean and land. Track steering and wind speed calculations used decay rates based on observed data and projected climate variables. Cyclone size was calculated dynamically, starting from LMI using a radial wind profile evolving along the track, capturing intensity-dependent size changes. Minimum pressure during the decay phase was modelled using a unified pressure-wind relationship, influenced by storm size and latitude. These components together generated synthetic cyclone datasets that replicated key physical and statistical characteristics of observed cyclones.

Only category 1 or greater cyclones making landfall in Africa were included in our final scenarios. Cyclones were assumed to begin at LMI, and tracks were terminated when modelled wind speed fell below tropical storm threshold (63 km h⁻¹). A counterfactual future scenario of cyclones reflecting present-day climate conditions was generated by resampling past cyclones, with Poisson rates of each category given by their frequency since 1980. Impact footprints were calculated based on R₁₈—the radius of damaging winds.

Parameterizing impact of extreme weather events on housing and interventions

Extreme weather events were modelled as disrupting access to three key suppressants of malaria transmission: (1) improved housing, (2) indoor vector control tools and (3) access to antimalarial treatment. The magnitude and duration of disruption was parameterized for each event class and severity on the basis of a mixed-methods approach: a literature review extracted 22 studies from the peer-reviewed and grey literature documenting the impact of extreme weather events (Extended Data Table 2). This process was augmented with 34 expert interviews. Parameters derived from this exercise were used to define recovery curves to be applied within footprints of extreme events, shown schematically in Extended Data Fig. 2, with parameters described in Extended Data Table 3. Acute and persisting impacts were differentiated, with the latter parameterized sigmoidally by time to 50% and 99% recovery. Uncertainty in the magnitude of disruption was then represented using a 50–150% scaling around the consensus central values.

Projecting impact of extreme weather events on housing and interventions

Disruptions to healthcare accessibility

Present-day access to effective treatment (EFT) for clinical malaria at location s was estimated using a composite of three surfaces generated by the Malaria Atlas Project⁴¹: probability of care-seeking c_s, use of antimalarial drugs ${p}_{s}^{{\rm{drug}}}$, and drug- and location-specific therapeutic efficacy ${E}_{s}^{{\rm{drug}}}$:

$${{\rm{EFT}}}_{s}={c}_{s}{\sum }_{{\rm{drug}}}{p}_{s}^{{\rm{drug}}}{E}_{s}^{{\rm{drug}}},$$

where ‘drug’ is one of artemisinin combination therapy (the first-line treatment for falciparum malaria in Africa), amodiaquine, sulfadoxine-pyrimethamine, chloroquine or quinine.

Using a database of health facility geolocations⁷³, a transport network model for Africa⁷⁴ and a least-cost-path journey time algorithm⁶⁶, we determined present-day (disruption free) travel time to health facilities for each grid cell. A bi-exponential relationship between these travel times ${{\mathcal{T}}}_{s}$ and 104,516 geo-located observations of propensity to seek care for fever was fitted, resulting in the function:

$${\rm{t}}\_{\rm{decay}}({{\mathcal{T}}}_{s})=0.1451\,\exp (-0.0798{{\mathcal{T}}}_{s})+0.5396\,\exp (-0.0008\,{{\mathcal{T}}}_{s}).$$

Health facilities located within the footprint of extreme weather extents were considered non-functional during acute impacts, with the proportion of facilities in each 5 × 5 km grid cell remaining non-functional in the post-acute period sampled from a binomial distribution, with probability of closure given by intersecting the appropriate recovery curve with a time-since-last-event surface. Given these dynamic functional facility geolocations, we recalculated travel time to health facilities for each scenario, by month, to 2050. In addition to facility closures, damage to road infrastructure was parameterized as travel time penalties derived from the recovery curves and time-since-event surfaces overlaid on the OpenStreetMap road network⁷⁴. Off-road travel time was recalculated by perturbing a friction surface⁷³ in the same way. The result was scenario-specific travel time to healthcare ${{\mathcal{T}}}_{s}^{{\rm{scenario}}}$. We then calculated care-seeking penalties relative to undisrupted conditions, so that at location s and time t:

$${{\rm{EFT}}}_{s,t}^{{\rm{scenario}}}=\frac{{\rm{t}}\_{\rm{decay}}({{\mathcal{T}}}_{s,\,t}^{{\rm{scenario}}})}{{\rm{t}}\_{\rm{decay}}({{\mathcal{T}}}_{s}^{{\rm{no}}\,{\rm{disruption}}})}{c}_{s}{\sum }_{{\rm{drug}}}{p}_{s}^{{\rm{drug}}}{E}_{s}^{{\rm{drug}}}.$$

Antimalarial efficacy, proportional usage of different antimalarials and secular variation in care-seeking behaviour were held constant, as was the undisrupted transport network.

Disruptions to ITN coverage and access to improved housing

ITN campaigns were simulated every three years on 1 January, commencing in 2025. As we aimed to model unmitigated impacts of extreme weather events, disrupted ITN coverage did not return to normal until the next simulated campaign. ITN access was assumed to be lost if access to housing was acutely disrupted. Access to improved housing was directly perturbed using the derived recovery curve.

Projecting ecological and disruptive effects of climate change

A set of scenarios was defined to derive the ecological, disruptive and combined impacts of climate change by the 2040s; these are described in Extended Data Table 1.

To model ecologically driven impacts of climate change we generated estimates of monthly PfPR_2–10 from 2019 to 2049 for each of the 14 ensemble members. A climate-change-free scenario (E₀) was defined as the ensemble mean of median PfPR_2–10, 2019–2022, and corresponding scenario of future changes in ecological suitability (E₁) as ensemble mean of median PfPR_2–10, 2040–2049. The ecological impact of climate change was thus estimated by the difference E₁ − E₀ (Fig. 1a).

We derived counterfactual and SSP 2-4.5 extreme weather event scenarios for a tractable subset of three GCMs: ACCESS-CM2, EC-Earth3-Veg-LR and MPI-ESM1-2-LR. For each ensemble member, two time-series of spatial PfPR_2–10 projections were calculated. Scenario D₀ was defined as the (three-GCM) ensemble mean of median PfPR_2–10, 2040–2049, generated with present-day TSI and HSI as in E₀ and with extreme events simulated as occurring at present-day frequency. Still holding TSI and HSI at present-day levels, SSP 2-4.5 projections of changing extreme weather event frequency were imposed to derive scenario D₁, so that the difference D₁ − D₀ isolated the impact of disruptive events due to climate change (Fig. 2a and Extended Data Table 1).

Ecological and disruptive impacts were combined by defining C₀ equal to D₀ (that is, present-day climate suitability and extreme event frequency). SSP 2-4.5 projections of changing climate suitability indices (as in E₁) and changing extreme weather frequency (as in D₁) were combined to generate C₁, the (three-GCM) ensemble mean of median PfPR_2–10, 2040–2049. The combined ecological and disruptive impacts were then calculated as C₁ − C₀ (Fig. 2b).

Projected PfPR_2–10 for each scenario was converted to clinical case incidence using an established natural history model⁴⁸. Gridded population projections consistent with SSP 2-4.5 were used to generate estimates of absolute cases⁴⁹. Projected time-series of clinical cases and mortality were scaled to align with 2022 official World Health Organization estimates, as reported in World Malaria Report 2023 (ref. ⁵⁰).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.