Data

Dengue case data

Annual reported dengue cases used in this study were compiled from 57 countries covering the period from 1980 to 2024 (Supplementary Fig. 13 and Supplementary Table 5). We mainly assembled this global dengue dataset from the OpenDengue website (https://opendengue.org/)⁶¹, with a time resolution set to “Year”. We also conducted an extensive search and cross-verified data across various sources, including the World Health Organization (WHO), the Pan-American Health Organization (PAHO), online databases, authorised Ministry of Health websites, established datasets⁵⁰, and published literature. For countries with missing data such as Thailand, Malaysia, Singapore, New Caledonia, Bhutan, China, and Maldives, we accessed their Ministry of Health or equivalent websites to fill in the gaps. Data from these sources were aggregated from weekly, monthly, and quarterly intervals to construct yearly datasets. Annual case counts for 2023 and 2024 were estimated based on the most complete annual cases for 2023 and 2024 as of the reporting date of 1st March 2025. Due to reporting delays, Anguilla, Bhutan, Dominica, and Venezuela were yet to finalise their case counts for 2024 by this date. We thus used a simple proportion of annual total cases by month model to predict their final 2024 annual case counts (Supplementary Text).

Note that data from 2020 to 2024 were not included in the main model training to avoid the impact of COVID-19. We included the recent years 2023 and 2024 in the sensitivity analysis to demonstrate the robustness of our main model. We also collected data from four temperate countries—Argentina, China, Chile, and the United States in our sensitivity analysis. Despite inter-country variations in case definitions and surveillance quality, intra-country variations over time can still be used to understand the drivers of dengue transmission.

ENSO indices, local climate, and population data

To assess the El Niño–Southern Oscillation (ENSO) events, we acquired the 1978–2024 time series of six ENSO indices (Supplementary Fig. 1a) from the National Climate Center (NCC) of China and the National Oceanic and Atmospheric Administration (NOAA) of the United States. Local climate observations were obtained from ERA5 monthly mean reanalysis product with a resolution of 0.25° × 0.25°, including air temperature at 2 m above the Earth’s surface and precipitation from 1980 to 2024⁶². The Gridded Population of the World (GPWv4) data at the 15 arc-minute resolution⁶³ was utilised to calculate the population-weighted climate for each country. We also collected country-level population statistics for the 1980–2024 time series from the World Population Prospects⁶⁴ for the offset term in our regression model.

Population-weighted climate calculation

We used population-weighted local climate to accurately capture the climate variations that affect people and dengue transmission. Population weighting gives more weight to areas with higher population density, ensuring that the climate conditions experienced by the majority of the population are better represented^31,65. This approach enhances the statistically valid assessment of climatic effects on dengue cases, allowing for more targeted public health interventions and improved resource allocation based on where people are most affected. Thus, based on gridded population data, we derived population-weighted averages of climatic variables:

where \({{{\mathrm{PrEnv}}}}_{i}\) represents the population-weighted climate of country i, \({N}_{g}\) and \({{Env}}_{g}\) denote the population and climate variables of across grid g, respectively, and n denotes the total number of grid cells in country i.

Correlation analysis

We calculated the Pearson correlation coefficient between ENSO indices and annual dengue cases. To ensure that the time series of ENSO indices precede the occurrence of the disease, we scrutinised the temporal correspondence for the north and south hemispheres. We conducted a linear detrend algorithm and z-score standardisation for ENSO indices and dengue cases to remove the effects of long-term warming trends.

We further assessed the collinearity between the annual changes in six ENSO indices based on the Pearson correlation coefficient and identified the NINO3.4 index as the most representative ENSO indicator. NINO3.4 is defined as the averaged anomaly of sea surface temperatures (SSTA) in the tropical Pacific Ocean over 5°S–5°N and 170°W–120°W. El Niño and La Niña events are characterised by a five-consecutive three-month running mean of the Niño 3.4 index exceeding +0.5 °C or falling below −0.5 °C, respectively^66,67. To focus on the season of peak ENSO activity, we used the average NINO3.4 over winter months for the following analysis between ENSO and dengue cases. Specifically, ENSO in a given year t was represented by the average NINO3.4 values for December of the preceding year t-1 and January and February of the current year t (referred to as DJF NINO3.4).

ENSO teleconnections with country-level climate

To match the country-level disease data resolution, we measured ENSO teleconnections as the extent to which each country’s climate is influenced by ENSO, accounting for the effects of temperature and precipitation, and different time scales at which teleconnections may manifest. Unlike previous approaches that examine the impacts of global climate and local climate on dengue epidemics separately, this method allows for a comprehensive study of the impact of ENSO-driven local climate on dengue epidemics, providing a more nuanced understanding of their relationships. Although ENSO may have opposing effects within large countries like Brazil⁶⁸, this study focuses on assessing the long-term effects of ENSO on country-level dengue cases due to data availability.

We used partial correlation to evaluate the teleconnections between interannual changes in DJF NINO3.4 and the local climatic variables for each country. First, we processed the 3-month running values of local climatic variables to eliminate random variations and ensure consideration of ENSO exposure over multiple months, including December_t-1–January_t–Februrary_t (DJF), January_t–Ferurary_t–March_t (JFM), Ferurary_t–March_t–April_t (FMA), and March_t–April_t–May_t (MAM). For each subset series of 3-month running mean temperature and cumulative precipitation during 1980–2019, we conducted a linear detrend algorithm and z-score standardisation to remove the effects of long-term warming trends and rare climatic events. Next, we calculated the partial correlation coefficients between these detrended and standardised time series of local climatic variables and DJF NINO3.4. Partial correlation was employed when analysing the ENSO–temperature correlation to control for precipitation, and vice versa, to account for the interactions between temperature and precipitation. This calculation yielded 4 correlation coefficients (DJF, JFM, FMA, MAM) for each country, separately for temperature and precipitation. We then identified the maximum absolute value of the correlation coefficients for each country. Finally, the positive and negative signs were reassigned to the maximum absolute values correspondingly to obtain the ENSO teleconnections with local temperature (\({\tau }^{{temp}}\)) and precipitation (\({\tau }^{{precip}}\)). The obtained ENSO teleconnections were then integrated into the ENSO-climate-dengue model. To assess the sensitivity of our findings to our chosen period over which ENSO teleconnections were calculated, we also calculated alternative teleconnection metrics over the summer, the whole year, and all the 3-month running values between December (t-1) and May (t). To address multiple hypotheses and control the false discovery rate, we employed the Benjamini-Hochberg (BH) method to assess the significance of the correlation coefficients⁶⁹.

ENSO–climate–dengue model

Our analysis aims to quantify the effects of ENSO on dengue epidemics, which requires us to distinguish ENSO from other confounding factors. We employed a distributed lag regression model using ordinary least squares estimation to estimate the effects of ENSO on dengue cases:

In Eq. 2, \({{{\mathrm{case}}}}_{{it}}\) denotes the reported cases in country i in year t. \({{{\mathrm{NINO}}}3.4}_{t}\) represents the NINO3.4 index in year t with the linear trend, while \({{{\mathrm{deNINO}}}3.4}_{t}\) represents the detrended NINO3.4 index in year t. Given that ENSO primarily drives local climate anomalies and extremes through oscillations^25,70, the detrended NINO3.4 index was used to unravel the effects of ENSO on dengue via local climate. This approach enables us to capture the oscillatory nature of ENSO while avoiding the repetition of multiple signals related to global warming in the equation. \({\tau }^{{{\mathrm{temp}}}}\) and \({\tau }^{{{\mathrm{precip}}}}\) allow us to isolate the impact of ENSO-driven temperature and precipitation on dengue epidemics from each other and seasonal averages. L denotes the lagged year in which the coefficient is estimated. Here, we estimated the impact of ENSO events in the current year, subsequent year and the year after that (i.e. a 2-year lag, j = 2) allowing us to capture longer-term indirect impacts of ENSO that may be triggered by ongoing epidemic momentum or longer-term environmental adaptations by the human and vector populations in response to the ENSO-induced extreme climate conditions^56,71,72. We tested the multicollinearity and confirmed low correlations among multiple lags and ENSO-driven climatic factors (Supplementary Table 6). \({{{\mathrm{population}}}}_{{it}}\) is the total population in country i in year t, serving as an offset term. \({\mu }_{i}\) is a time-invariant country fixed effect, which controls for differences among countries’ ability to control, detect and report dengue cases. We did not include year fixed effects in our main analysis due to their high collinearity with the NINO3.4 index, but we tested the influence of important year-varying confounders that may affect dengue cases (i.e., the changes in WHO case definition in 2009 and yearly variations in the circulating serotype in 9 countries from 1990 to 2019 based on the dengue sequences reported in the Bacterial and Viral Bioinformatics Resource Center⁷³).

Our analysis focused on the cumulative coefficients \({\Omega }_{{iL}}^{{{\mathrm{temp}}}}\) and \({\Omega }_{{iL}}^{{{\mathrm{precip}}}}\), representing the accumulated effects of ENSO on reported dengue cases for each country i in L years after the event, via temperature and precipitation, respectively:

These influences differ between countries and lag lengths, depending on the strength of the coupling between each country’s temperature and precipitation with ENSO. The inclusion of lags from year L to year j helps differentiate between transient and persistent impacts of ENSO on dengue cases. If the coefficient is significantly different from zero (P ≤ 0.05), it suggests that ENSO has a persistent impact on dengue cases. Conversely, if it is not significantly different from zero, we cannot reject the hypothesis and conclude that ENSO only affects transient dengue outbreaks. Confidence intervals were estimated using 1000 bootstrap samples. We also estimated the confidence intervals using a Bayesian generalised linear model with 2000 bootstrap samples, and obtained similar results (Supplementary Table 7).

Despite the inclusion of lagged terms, the Ljung-Box test (Q* = 2465.8, P ≤ 0.01) and the Breusch-Pagan test (BP = 46.3, P ≤ 0.01) indicated that temporal autocorrelation and heteroscedasticity still exist in our model. To avoid the impact of temporal autocorrelation, we applied the Newey-West adjustment⁷⁴ to recalculate standard errors of the estimates. The Newey-West adjustment has been commonly used in econometric analysis and epidemiological studies⁷⁵ to ensure the validity of statistical inference. The adjusted standard errors validated that temporal autocorrelation does not affect the significance of our regression coefficients (Supplementary Table 1). We also included case_t-1 as an explanatory variable to control the autoregressive effect and validated that our regression results remain robust.

Magnitude of dengue epidemics from historical extreme El Niño events

The regression coefficients derived from Eq. 2, \({\theta }_{L}^{{temp}}\) and \({\theta }_{L}^{{precip}}\), not only provide estimates of the change in dengue cases per one unit change in the NINO3.4 index, but can also be applied to historical NINO3.4 values to estimate the change in cases due to El Niño events. Here, we focused on the four strongest El Niño events of 1982–83, 1997–98, 2015–16, and 2023–24. We defined counterfactual ENSO wherein El Niño and La Niña events did not occur by setting the corresponding positive and negative NINO3.4 values to zero. We then used the Delta method^31,76 to calculate the case change rate [\(\Delta {{case}}_{i(t+L)}\)] in the occurrence year (L = 0), following year (L = 1), and the second year (L = 2) due to the El Niño event via temperature and precipitation:

where \({{deNINO}3.4}_{t}\) represents the observed NINO3.4 value in the year of the El Niño event (t), and \({{{\mathrm{deNINO}}}3.4}_{t}^{0}\) represents the counterfactual ENSO (zero) in that year. We added this case change rate to the observed cases in those three lagged years (L = 0, 1, 2) for each country, obtaining the change from counterfactual (without the event) to observed (with the event) cases. The cumulative effects of each El Niño event were calculated as the sum of changes in three lagged years after the event. Global case changes due to the event were then calculated as the sum of the changes in dengue cases in 57 countries. Note that our estimate of the global change in cases attributable to the 2023–24 El Niño was based solely on the observed cases in 2024, the year of occurrence, and that we would expect this estimate to increase when dengue case data for 2025 and 2026 become available.

CMIP model selection

We used multiple ensemble members of CMIP6 climate models that are available under four emission scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) to predict changes in dengue cases due to ENSO evolution under future warming. We tested 88 multi-model ensemble members for SSP1-2.6, 150 for SSP2-4.5, 134 for SSP3-7.0, and 94 for SSP5-8.5. We measured ENSO amplitude as the standard deviation of the NINO3.4 time series⁴⁵. We also assessed the frequency of ENSO events but found a high uncertainty with great variations across multi-model ensemble members and a large bias between simulated and observed frequencies of ENSO events during 1980–2019. Thus, we chose the relatively robust ENSO amplitude to predict dengue cases.

To ensure that our projections are feasible, we defined “skilful” ensemble members as the absolute value of the bias between simulated and observed ENSO amplitude during 1980–2019 is less than 50% of the observed ENSO amplitude⁷⁷. Based on those selected skilful ensemble members, we calculated the population-weighted monthly temperature and daily precipitation to assess the teleconnections for each country during 1940–2019 and 2020–2099. The change rates in ENSO amplitude and teleconnection strength (the absolute value of teleconnection) were obtained to project changes in dengue cases in the next section.

Projecting future effects of ENSO

To estimate the impact of projected future changes in ENSO on dengue cases, we compared scenarios where ENSO is projected to change with a counterfactual where it remains the same as over the observed period (1980–2019). We defined the observed ENSO and teleconnections as the “counterfactual future”, and the “future” ENSO and teleconnections were calculated by adding the change rate between the historical and future simulations to the observed values (Eqs. 6–7). This method ensures that the “counterfactual future” retains the same amplitude and teleconnections as the observed ENSO from the historical period. As such, the change in cases between the “counterfactual future” and “future” represents the dengue cases associated with projected changes in ENSO evolution under different warming scenarios in the 21st century^31,78.

We then estimated the change rate in dengue cases due to future ENSO changes in the scenarios of interest (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5) as

where \({{deNINO}3.4}_{F}^{e}\) and \({{deNINO}3.4}_{{CF}}^{e}\) represent the future and counterfactual future ENSO time series, respectively. \({\tau }_{i,F}^{c}\) and \({\tau }_{i,{{\mathrm{CF}}}}^{c}\) represent the future and counterfactual future teleconnections for each country i, respectively. c denotes the local climatic variables. We projected the country-level change rate in dengue cases for multi-model ensemble members, using 1000 bootstrap estimates of our empirical model. The average change rate in global dengue cases due to changes in ENSO amplitude and teleconnections between future and counterfactual future was calculated as the mean of country-level change rates. We also estimated the average change rate in global dengue cases due to changes in El Niño and La Niña events between future and counterfactual future, based on the average detrended NINO3.4 value during the past El Niño and La Niña events separately, where e denotes the ENSO events.

Uncertainty in the nonlinear association between future warming and cases

Given that the optimal temperature for dengue transmission is estimated to be around 29 °C⁴⁸, we expected that our predictions might be less powerful for the countries with temperatures exceeding 29 °C during 2020–2099. Here, we used the monthly average temperature for each country during 2020–2099 to assess the power of our predictions of dengue risk based on our linear regression model.

We first corrected the CMIP-predicted monthly average temperatures from 2020 to 2099, based on the bias between the observed temperatures and CMIP-predicted temperature for each country and month over 1980–2019:

where \({T}_{{{\mathrm{imt}}}}\) represents the corrected monthly average temperature, \({T}_{{{\mathrm{imt}}}}^{{{\mathrm{CMIP}}}}\) represents the monthly average temperature predicted from each CMIP6 simulation in country i in month m and year t, \({\widetilde{T}}_{{{\mathrm{im}}}}^{{{\mathrm{observed}}}}\) denotes the observed temperature in country i in month m over 1980–2019, and \({\widetilde{T}}_{{{\mathrm{im}}}}^{{{\mathrm{CMIP}}}}\) denotes the temperature predicted from this CMIP model over the same period. This correction largely reduced the Root Mean Squared Error (RMSE) from ~4.5 to ~3.5 between observed and CMIP-predicted temperatures over 1980–2019. We then calculated the proportion of countries with average summertime temperatures and annual maximum temperatures during 2020–2099 exceeding 29 °C to represent the uncertainty in our projections of rising dengue risk.

Reporting summary

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