Associations between weather and Plasmodium vivax malaria in an Amazonian elimination setting: a distributed lag analysis from 2017 to 2024 Authors: Gabriella Barratt Heitmann, Xue Wu, Anna T. Nguyen, Astrid Altamirano-Quiroz, Sydney R. Fine, Bryan Fernandez-Camacho, Antony Barja, Renato Cava, Verónica Soto-Calle, Hugo Rodriguez, Gabriel Carrasco-Escobar, Adam Bennett, Alejandro Llanos-Cuentas, Erin A. Mordecai, Michelle S. Hsiang, Jade Benjamin-Chung Published: 2026 ABSTRACT BackgroundThere is limited evidence regarding the association between weather and Plasmodium vivax (Pv), particulary in Latin America where Pv is the predominant malaria species and key challenge for countries to achieve malaria elimination.MethodsWe analyzed the association between weather and Pv malaria incidence from 2017 to 2024 in 136 communities in the Peruvian Amazon. Monthly community-level incidence was calculated using Pv case data from Notiweb, the national epidemiological surveillance system, and population census data. Predictors included weekly minimum and maximum temperature and total weekly precipitation and were calculated using hourly weather from the climate dataset ERA5. Non-linear distributed lag models were fit using a lookback period of 2–16 weeks. Temperature models were adjusted for total precipitation; precipitation models were adjusted for maximum temperature. Sub-group analyses were conducted by community type (adjacent to river versus highway) and El Niño Southern Oscillation (ENSO) period.ResultsMinimum temperature at the 90th percentile (23.7°C) was associated with 10% (95% CI 5–14%) higher malaria incidence compared to the 5th percentile (20.5°C) at a 7-week lag. Maximum temperature at the 90th percentile (33.7°C) was associated with 10% (95% CI 8–13%) higher malaria incidence compared to the 5th percentile (29.6°C) at a 9-week lag. Total weekly precipitation at the 90th percentile (1000 mm) was associated with 29% (95% CI 24–33%) higher malaria incidence compared to weeks with the 5th percentile (57 mm) at an 11-week lag. Incidence was higher and associations were stronger in communities adjacent to rivers versus highways. Malaria incidence was lower during El Niño periods, and there was evidence of interaction on the multiplicative scale for the association between incidence, all weather predictors, and ENSO period.ConclusionsPv malaria incidence was positively associated with higher temperatures and precipitation in an elimination setting in Peru, particularly in riverine communities during non-El Niño years, with longer lag periods than previously reported for such associations. These findings can inform malaria elimination interventions to combat the long-lasting effects of weather on Pv transmission.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12936-026-06010-y. Background As countries approach malaria elimination, the proportion of malaria infections due to Plasmodium vivax (Pv) increases [1]. Pv presents unique challenges for malaria elimination efforts because of its ability to survive in wide ranging environments, including temperate settings, as a result of its dormant hypnozoite stage, which re-emerges to cause relapse infections [2]. Pv only invades reticulocytes, early stage red blood cells, and infections are low density and often missed by standard diagnostics. Further, a high proportion of infections are asymptomatic, and thus evade detection through standard surveillance. These characteristics make Pv difficult to detect, treat, and ultimately eliminate. The relationship between weather and Plasmodium falciparum (Pf) malaria transmission is well-documented, leading to routine deployment of seasonal malaria interventions, such as seasonal malaria chemoprevention [3], in places where Pf is endemic. However, the relationship between weather and Pv malaria transmission, whether seeded by relapses or new infections, is not as well understood, preventing malaria programs from considering seasonal targeting of interventions to optimize their impact. However, there is strong biologic plausibility for temperature and precipitation to influence Pv transmission via similar mechanisms to Pf (Fig. 1) [4–10]. Precipitation can increase standing water presence, increasing breeding ground for mosquitoes [11]; temperature has strong effects on the lifecycle of the Anopheles vector, including larval development, biting rate, parasite incubation rate, and longevity [12, 13]. Further, there is emerging evidence that weather could directly trigger relapse Pv infections [14], and weather may indirectly drive Pv infections via the saliva of uninfected mosquitoes, which contain antigens that may trigger relapse (biting rate is temperature-sensitive) [15].Fig. 1Hypothesized mechanisms of the impact of temperature and precipitation on Plasmodium vivax infection Hypothesized mechanisms of the impact of temperature and precipitation on Plasmodium vivax infection Further, if seasonal patterns for Pv exist, strong spatial heterogeneity [16, 17] in endemic regions requires local characterization to inform intervention targeting. Over 90% of Latin America’s malaria burden is concentrated in the Amazon Basin [18]. Parasite populations and vector behavior vary significantly in this region [19–21], and this variation is likely driven by adaptations to different environmental conditions [20, 22, 23]. Prior studies of the relationship between weather and malaria incidence in the Amazon have not focused exlcusively on Pv even though Pv is estimated to account for 72% of the transmission burden [24]. Finally, the Amazon Basin is undergoing rapid environmental and climatic changes, including deforestation, warming, and drought [25–27], that may alter transmission dynamics at a fine scale, further emphasizing the need for local characterization. Our objective was to assess the influence of rainfall and temperature on Pv incidence in Loreto Region, Peru, where the maximum temperatures fall between 30 and 35°C, which are above the estimated thermal optimum for malaria transmission (25°C) [28]. Daily temperature extremes are thought to limit transmission via detrimental impacts on parasite development, adult mosquito survival, egg survival, and mosquito development [28, 13, 29]. However, to our knowledge, the thermal biology of Pv transmission by Nyssorhynchus (formerly Anopheles) darlingi, the primary vector in this region, has also not been well characterized by laboratory thermal biology studies. Information about climatic drivers of Pv transmission is needed to inform effective tailoring and targeting of interventions for malaria elimination in this region. Methods Study design We analyzed malaria incidence data collected from January 2017 to December 2024 in all 136 communities within 8 h of Iquitos, the periurban center of Loreto Region, Peru. This area covering 2,569 km2 includes the districts of San Juan Bautista, Punchana, and Alto Nanay (Fig. 2). Farming, logging, and fishing are common occupations in the study area, and the mobility patterns associated with logging and fishing may increase malaria transmission and case importation from other areas [30, 31]. Our study area includes more densely populated, peri-urban communities concentrated around the Iquitos-Nauta highway, and more isolated communities concentrated around tributaries running from Iquitos deep into the Amazon rainforest. The Peruvian Ministry of Health (MINSA) conducted a census in 2017 that was used to estimate the starting population size of each community. The average population size of each community was 321 individuals (range: 15–3,223).Fig. 2Community-level Plasmodium vivax malaria incidence within the study area. Dark green areas indicate forest cells from MapBiomas. Circles with black outlines represent the 136 study communities, colored by malaria incidence per 10,000 person-weeks over the study period. Community outlines represent type: squares for highway-adjacent communities and circles for riverine. Community-level Plasmodium vivax malaria incidence within the study area. Dark green areas indicate forest cells from MapBiomas. Circles with black outlines represent the 136 study communities, colored by malaria incidence per 10,000 person-weeks over the study period. Community outlines represent type: squares for highway-adjacent communities and circles for riverine. The Loreto Region in Peru includes a large swath of the Peruvian Amazon [32]. Since the 1990 s, rapid urbanization and deforestation in the area surrounding Iquitos, the capital of the Department of Loreto, has contributed to increased malaria burden [32], likely through increased forest edge habitat, which promotes mosquito breeding, survival, and biting [25]. While this region is approaching elimination, sustained, endemic transmission remains, with an annual incidence rate of 17.4 cases per 1000 in 2019 (annual incidence < 100 per 1000 is considered very low transmission [33]) [34], and Pv malaria accounted for approximately 81% of the malaria burden in 2024 [35]. Transmission typically peaks in Loreto between February and July [32] and there is a sizable burden of asymptomatic infections [6]. Asymptomatic infections typically go undetected and uncounted in routine malaria surveillance in Peru [6], since only febrile individuals are tested. From 2005 to 2010, there was a large effort towards malaria elimination in Peru via the Project for Malaria Control in Andean Border Areas (PAMAFRO) program that improved case management, including through deployment of community case workers and distribution of insecticide-treated bed nets [32]. Transmission declined during the project, but the project did not achieve elimination, and transmission has since risen [36]. Following PAMAFRO, the government of Peru adopted Plan Malaria Cero (PMC) in 2017 to achieve malaria elimination by 2030. Activities conducted over the study period as part of Plan Malaria Cero include strong case management of suspected malaria cases presenting to health facilities, distribution of long-lasting insecticide-treated bed nets (ITN), and outbreak responsive activities such as active case detection (ACD), fogging, and indoor residual spraying [37]. Outcome data The primary outcome was weekly community-level Pv malaria incidence measured in surveillance data. Data primarily included passive case detection, though reactive and active case detection also occurred over the study period and were recorded in surveillance data. Both physical records stored in health posts and online records compiled by MINSA were used in this analysis. We cross-referenced physical records with online records for names and infection dates to remove any duplicates. Febrile patients presented for care at a health post and were tested for malaria via microscopy. According to PMC policy [37], blood smears are stained with 2% Giemsa for 30 minutes. Parasite densities are calculated from the number of asexual parasites per 200 leukocytes (or per 500, if < 10 asexual parasites/200 leukocytes), assuming a leukocyte count of 8,000/µL. A blood smear is considered negative if examination of 100 high power fields does not reveal asexual parasites. Thin smears are used for parasite species identification. Slides are read by two microscopists. If there are discordant results, the results are determined by a third microscopist. Only Pv cases were included in the present study, and we were unable to distinguish between primary and relapse infections from microscopy data. Cases were matched to community census data to confirm community of residence and then matched to geocoordinates for the community centroid. Population counts per community were calculated from a government census conducted from 2017. The FocaL Mass Drug Administration for Vivax Malaria Elimination (FLAME) [38] trial in Peru (NCT05690841) conducted a census on a subset of 30 communities in 2023. We estimated adjusted population counts per year by calculating the annual growth rate using the 2023 and 2017 census counts in the subset. The same flat annual growth rate was applied to all 136 communities in the study area with a population greater than 15 people. We obtained data for ITN distribution from MINSA with delivery dates for study communities from 2018 to 2024. Time since most recent ITN delivery was discretized into never received, 0–1 year, 1–3 years, and more than 3 years. Environmental variables We obtained temperature and precipitation data from the ERA5-Land Hourly dataset, 11 km resolution, collected by the Copernicus Climate Change Service [39]. We used the air temperature at 2 meters above the surface (K) band for all temperature variables and the total precipitation (m) band for precipitation variables. We matched weather variables to incidence data via community centroid. There was no missing weather data over the study period. We aggregated hourly temperature values into the weekly minimum and weekly maximum temperature; and total precipitation observed each week. All data extraction was performed using the Python API for Google Earth Engine on Google Colab servers. Aggregation over time and statistical modeling were performed in R version 4.2.1. Furthermore, the Amazon region is subject to longer-term climate oscillations due to the El Niño Southern Oscillation (ENSO). The ENSO cycle is characterized by hot, dry conditions during El Niño years and rainy conditions during La Niña years; these periods oscillate with neutral periods every 3–7 years [40]. We obtained data on the Oceanic Niño Index (ONI), a measure of the strength of the El Niño-Southern Oscillation (ENSO), from the National Oceanic and Atmospheric Association National Weather Service Climate Prediction Center. A 1-unit increase in ONI represents a 1°C higher average sea-surface temperature compared to a 30-year reference period [40]. We manually designated riverine versus highway communities via Google Earth hybrid satellite image using QGIS version 3.32. We assigned communities with no land connections to the Iquitos-Nauta highway as riverine, and all other communities as highway. We checked our designations with ground-truth observations from 2023 for the FLAME community subset. Statistical analysis We published a pre-analysis plan at https://osf.io/f89te. Deviations from the pre-analysis plan are listed in Table S1. We removed years 2020 and 2021 from our primary analysis because of strong changes in incidence and case detection resulting from behavior changes and health system capacity during the COVID-19 pandemic. To assess the timing of the association between malaria incidence and weather variables, we we fit distributed lag models. We chose a 2–16 week lookback period (i.e., an infection during epidemiologic week 20 looked back to weeks 4, 5, …, 18) to account for variation in the potential cumulative effect of rainfall and temperature on Pv transmission and long-lasting impacts on relapse cases (Fig. 1). We set the upper limit at 16 weeks (4 months) to ensure we captured at least one relapse cycle in accordance with studies of tropical strains of Pv documenting median time to first relapse 6–9 weeks after an initial infection [4] and frequent relapses 3–4 weeks apart for up to 4 months after the initial infection [9]. We were unable to adjust for season over our study period, as data-based definitions (i.e., consecutive weeks with rolling average of precipitation above a percentile cutoff) failed to delineate clear seasons. Given prior literature states seasons generally last for 6 months [32], constraining our lookback period to 4 months also limited the effects of seasonal confounding. For weather exposures (weekly minimum temperature, weekly maximum temperature, and total precipitation), we fit non-linear distributed lag (DL) models [41] with a log link and Poisson family using weekly malaria case counts per community as the dependent variable with an offset for log community population size using the R package DLNM version 2.4.7 [42]. For the distributed lag cross-basis function, we specified natural splines with 2 df for the exposure–response dimensions, and 1 internal lag knot to allow for differences in early versus later lag periods. In model testing, this combination best balanced flexibility with stability compared to specifications with higher values for the natural spline degrees of freedom and lag knots, which showed evidence of overfitting. In all models, we restricted weather measurements in the input data to observations between the 5th and 95th percentile of the variable’s empirical distribution to remove outlier effects. All incidence ratios were calculated relative to reference weeks. For all models, we used the 5th percentile of the data range as the reference value: 20.5°C for minimum temperature, 29.6°C for maximum temperature, and 57 mm of precipitation. Temperature variables were adjusted for the cross-basis matrix for total precipitation, and total precipitation was adjusted for the cross-basis matrix for minimum temperature, in accordance with AIC/BIC model testing. All DL models included fixed effects for year, community, and years since ITN distribution, and a smooth term for ONI with 3 knots. All positive cases were matched to weather data. Thus, we performed a complete case analysis. Because incidence data for the weather models was aggregated at the community level, we did not test for spatial autocorrelation or adjust for spatial clustering. To assess the relationship between each predictor value and malaria incidence, we fit generalized additive models (GAMs) with a Poisson family and log link and an offset for community population size. We fit models for the week with the strongest association in the DL model (primary lag). All GAMs included fixed effects for year, community, and years since ITN delivery and a smooth term for ONI with 3 knots. We derived confidence intervals from the 2.5th and 97.5th percentiles of a 1,000-replicate bootstrap distribution. For each bootstrap replicate, we used g-computation to fix the predictor at each point across its observed range and average model predictions across all observed covariate combinations. We conducted subgroup analyses using the primary lag for temperature and precipitation predictors stratifying by ENSO periods and community type. We defined periods of 5 or more consecutive months of sea surface temperature abnormality of 0.5°C or above/−0.5°C or below as El Niño/La Niña periods, respectively [40]. To assess interaction effects between weather and ENSO period or community type on the multiplicative scale, we fit generalized linear models (GLMs) that included the same covariates from GAM models as fixed effects. Results During the study period from January 2017 to December 2024, minimum temperature varied from 17 to 25°C and maximum temperature varied from 27 to 43°C across all study communities (Fig. 3A). Total weekly precipitation varied from 7 mm per week to 2,587 mm per week across all study communities (Fig. 3B). There were 112 El Niño weeks, 158 La Niña weeks, and 147 neutral weeks over the study duration, and incidence varied by ENSO period, with the highest incidence during neutral periods. Malaria incidence was highest from 2017 to 2019, then decreased dramatically in 2020 and 2021 and increased to levels below the 2019 levels in 2022–2024. Malaria incidence did not appear to follow a distinct seasonal trend (Fig. 3C). The mean incidence for all communities was 8.5 cases per 10,000 person-weeks.Fig. 3Weather and Plasmodium vivax malaria incidence trends over the study period. All weather variables are aggregated into weekly means for all study communities. Incidence is aggregated monthly. Shaded blue regions represent La Niña periods, while shaded beige regions represent El Niño periods. In plot A, the red line tracks the average of the maximum temperatures observed each week in all study communities, and the orange line tracks the average minimum temperature observed each week Weather and Plasmodium vivax malaria incidence trends over the study period. All weather variables are aggregated into weekly means for all study communities. Incidence is aggregated monthly. Shaded blue regions represent La Niña periods, while shaded beige regions represent El Niño periods. In plot A, the red line tracks the average of the maximum temperatures observed each week in all study communities, and the orange line tracks the average minimum temperature observed each week Temperature There was a positive association between minimum temperature and malaria incidence at 2–13 weeks lags, and a negative association at 14–16 week lags when comparing the 90th percentile of minimum temperature (23.7°C) to the 5th percentile (20.5°C) (Fig. 4A, C). Where this association was highest, at a 7-week lag, the association between minimum temperature and incidence was positive and nonlinear (Fig. 4B). At minimum temperatures of 20.5–22°C, incidences were similar. At higher temperatures, there was a positive, non-linear association with estimated incidence of 18 cases per 10,000 at 20.5°C compared to 30 cases per 10,000 at 23.8°C. Compared to weeks with the lowest minimum temperature, weeks with a minimum temperature of the 90th percentile were associated with up to 10% higher malaria incidence (95% CI 6–14%), or incidence ratio of 1.10 at a 7-week lag (Figure 4A).Fig. 4Associations between Plasmodium vivax malaria incidence, temperature, and precipitation over different lag periods. The associations were fit using distributed lag non-linear models. Plots A–C show the association for minimum weekly temperature, and plots D–F show the association for maximum weekly temperature, and plots G–I show the association for weekly total precipitation. Plots A, D, and G show the association between temperature and malaria incidence at the 90th percentile predictor value compared to the reference. For minimum temperature, the 90th percentile is 23.7°C and the reference lowest minimum temperature is 20.5°C. For maximum temperature, the 90th percentile is 33.7°C and the reference lowest maximum temperature is 29.6°C. For total precipitation, the 90th percentile is 1,000 mm and the reference is 57 mm. Temperature models adjusted for total precipitation; total precipitation models adjusted for minimum temperature. All models were adjusted for community ID, year, years since insecticide-treated bednet delivery, and a smooth for Oceanic Niño Index. Incidence is presented per 10,000 person-weeks. Associations in plots B, E, and F were fit for the most significant lag from plots A, D, and G, respectively, using generalized additive models (GAMs). Confidence intervals were calculated using 1,000 bootstrap iterations with g-computation. GAMs do not account for temporal autocorrelation between weeks Associations between Plasmodium vivax malaria incidence, temperature, and precipitation over different lag periods. The associations were fit using distributed lag non-linear models. Plots A–C show the association for minimum weekly temperature, and plots D–F show the association for maximum weekly temperature, and plots G–I show the association for weekly total precipitation. Plots A, D, and G show the association between temperature and malaria incidence at the 90th percentile predictor value compared to the reference. For minimum temperature, the 90th percentile is 23.7°C and the reference lowest minimum temperature is 20.5°C. For maximum temperature, the 90th percentile is 33.7°C and the reference lowest maximum temperature is 29.6°C. For total precipitation, the 90th percentile is 1,000 mm and the reference is 57 mm. Temperature models adjusted for total precipitation; total precipitation models adjusted for minimum temperature. All models were adjusted for community ID, year, years since insecticide-treated bednet delivery, and a smooth for Oceanic Niño Index. Incidence is presented per 10,000 person-weeks. Associations in plots B, E, and F were fit for the most significant lag from plots A, D, and G, respectively, using generalized additive models (GAMs). Confidence intervals were calculated using 1,000 bootstrap iterations with g-computation. GAMs do not account for temporal autocorrelation between weeks The association between maximum temperature and malaria incidence was non-linear and hump-shaped across lag time; there was a positive association between maximum temperature and malaria incidence at 4–14 week lags when comparing the 90th percentile of maximum temperature (33.7°C) to the 5th percentile (29.6°C) (Fig. 4D, F). Where this association was highest, at a 9-week lag, incidence increased approximately linearly from an estimated 16 cases per 10,000 to 26 cases per 10,000 as maximum temperature increased from 29.6 to 34°C (Fig. 4E). Compared to weeks with the lowest maximum temperature, weeks with a maximum temperature of the 90th percentile were associated with up to 12% higher malaria incidence (95% CI 9–15%), or incidence ratio of 1.12 at a 9-week lag (Fig. 4D). Precipitation The association between precipitation and malaria incidence was non-linear and positive at lags of 5–16 weeks, when comparing the 90th percentile of precipitation (1000 mm) to the 5th percentile (57 mm) (Fig. 4G, I). Where this association was highest, at an 11-week lag, incidence increased approximately linearly from an estimated 18 cases per 10,000 to 24 cases per 10,000 as precipitation increased from 57 to 1161 mm (Fig. 4H). At an 11-week lag, compared to weeks with no precipitation, weeks with precipitation at the 90th percentile were associated with up to 25% higher malaria incidence (95 CI 21–28%), or incidence ratio of 1.25 (Fig. 4G). Community type sub-group Malaria incidence was higher in riverine versus highway-adjacent communities (Table 1, Fig. 5A–C). For minimum temperature, GLMs showed that associations were stronger in riverine versus highway-adjacent communities: the incidence ratio for a 1°C increase in minimum temperature was 1.27 (95% CI 1.23–1.31) in riverine communities compared to 1.10 (1.02–1.19) in highway communities (interaction p = 0.001) (Table 1). For maximum temperature, GLMs showed that associations were stronger in riverine versus highway-adjacent communities: the incidence ratio for a 1°C increase in maximum temperature was 1.13 (95% CI 1.11–1.15) in riverine communities compared to 1.03 (0.98–1.09) in highway communities (interaction p = 0.001) (Table 1). For precipitation, GLMs showed associations were weaker for riverine versus highway-adjacent communities: the incidence ratio for a 500 mm increase in precipitation was 1.04 (95% CI 1.00–1.08) in riverine communities compared to 1.59 (95% CI 1.44–1.76) in highway communities (interaction p < 0.001) (Table 1). Table 1Plasmodium vivax malaria incidence ratios across temperature and precipitation, by community typeCommunity typePredictor rangeIncidence rangeIncidence ratio (95% CI)Interaction p-valueMinimum temperature Highway20.5–23.97.3–13.51.10 (1.02–1.19)– Riverine20.4–23.817.8–31.41.27 (1.23–1.31)0.001Maximum temperature Highway29.5–33.95.2–10.81.03 (0.98–1.09)– Riverine29.6–34.017.4–27.11.13 (1.11–1.15)0.001Precipitation Highway51–11397.7–16.31.59 (1.44–1.76)– Riverine57–116118.3–23.01.04 (1.00–1.08) < 0.001We fit generalized linear models (GLMs) with interaction terms to obtain incidence ratios across temperatures and precipitation values. Treating highway communities as the reference, p values for interaction are shown. Temperature GLMs adjusted for precipitation at a concurrent lag, while precipitation models adjusted for minimum temperature at the concurrent lag. All GLMs were adjusted for community ID, year, Oceanic Niño Index, and years since insecticide-treated bednet delivery. All temperature ranges are reported in °C. Precipitation ranges are reported in millimeters. Incidence ranges are reported per 10,000 person-weeks. Incidence ratios are reported per 1°C increase for temperature predictors and per 500 mm increase for precipitation.Fig. 5Plasmodium vivax malaria incidence by temperature and precipitation within El Niño-Southern Oscillation (ENSO) period and community type subgroups. All weather values represent weekly values. Associations were fit for the strongest lag identified in the distributed lag model. Community type was determined by proximity to the Iquitos-Nauta highway or a tributary to the Amazon River. ENSO periods were defined according to the Oceanic Niño Index (ONI) values for sea surface temperature during the study period. Periods with 5 + consecutive months of ONI ≥ 0.5 were defined as El Niño periods. Periods with 5 + consecutive months of ONI ≤ −0.5 were defined as La Niña periods. All other periods were considered neutral periods. Associations were fit for the most significant lag identified in Figure 4 using generalized additive models (GAMs). Confidence intervals were calculated using 1,000 bootstrap iterations with g-computation. GAMs do not account for temporal autocorrelations. Table 1Plasmodium vivax malaria incidence ratios across temperature and precipitation, by community typeCommunity typePredictor rangeIncidence rangeIncidence ratio (95% CI)Interaction p-valueMinimum temperature Highway20.5–23.97.3–13.51.10 (1.02–1.19)– Riverine20.4–23.817.8–31.41.27 (1.23–1.31)0.001Maximum temperature Highway29.5–33.95.2–10.81.03 (0.98–1.09)– Riverine29.6–34.017.4–27.11.13 (1.11–1.15)0.001Precipitation Highway51–11397.7–16.31.59 (1.44–1.76)– Riverine57–116118.3–23.01.04 (1.00–1.08) < 0.001We fit generalized linear models (GLMs) with interaction terms to obtain incidence ratios across temperatures and precipitation values. Treating highway communities as the reference, p values for interaction are shown. Temperature GLMs adjusted for precipitation at a concurrent lag, while precipitation models adjusted for minimum temperature at the concurrent lag. All GLMs were adjusted for community ID, year, Oceanic Niño Index, and years since insecticide-treated bednet delivery. All temperature ranges are reported in °C. Precipitation ranges are reported in millimeters. Incidence ranges are reported per 10,000 person-weeks. Incidence ratios are reported per 1°C increase for temperature predictors and per 500 mm increase for precipitation. Plasmodium vivax malaria incidence ratios across temperature and precipitation, by community type We fit generalized linear models (GLMs) with interaction terms to obtain incidence ratios across temperatures and precipitation values. Treating highway communities as the reference, p values for interaction are shown. Temperature GLMs adjusted for precipitation at a concurrent lag, while precipitation models adjusted for minimum temperature at the concurrent lag. All GLMs were adjusted for community ID, year, Oceanic Niño Index, and years since insecticide-treated bednet delivery. All temperature ranges are reported in °C. Precipitation ranges are reported in millimeters. Incidence ranges are reported per 10,000 person-weeks. Incidence ratios are reported per 1°C increase for temperature predictors and per 500 mm increase for precipitation. Plasmodium vivax malaria incidence by temperature and precipitation within El Niño-Southern Oscillation (ENSO) period and community type subgroups. All weather values represent weekly values. Associations were fit for the strongest lag identified in the distributed lag model. Community type was determined by proximity to the Iquitos-Nauta highway or a tributary to the Amazon River. ENSO periods were defined according to the Oceanic Niño Index (ONI) values for sea surface temperature during the study period. Periods with 5 + consecutive months of ONI ≥ 0.5 were defined as El Niño periods. Periods with 5 + consecutive months of ONI ≤ −0.5 were defined as La Niña periods. All other periods were considered neutral periods. Associations were fit for the most significant lag identified in Figure 4 using generalized additive models (GAMs). Confidence intervals were calculated using 1,000 bootstrap iterations with g-computation. GAMs do not account for temporal autocorrelations. ENSO sub-group Malaria incidence was highest during neutral periods and lowest during El Niño periods (Table 2, Fig. 5D–F). Using GLMs, associations between minimum temperature and malaria incidence were positive when stratifying by ENSO period and strongest during neutral and El Niño periods. The incidence ratio for a 1°C increase in minimum temperature was 1.34 (95% CI 1.28–1.39) during neutral periods, 1.28 (95% CI 1.21–1.36) during El Niño periods (interaction p = 0.269), and 1.04 (95% CI 0.99–1.10) during La Niña periods (interaction p < 0.001) (Table 2). Table 2Plasmodium vivax malaria incidence ratios for malaria incidence across temperature and precipitation, by ENSO (El Niño-Southern Oscillation) period.ENSO periodPredictor rangeIncidence rangeIncidence ratio (95% CI)Interaction p-valueMinimum temperature Normal19.9–23.717.3–31.71.34 (1.28–1.39)– El Niño21.4–24.04.3–16.71.28 (1.21–1.36)0.269 La Niña20.2–23.414.2–20.21.04 (0.99–1.10) < 0.001Maximum temperature Normal29.5–33.720.4–25.91.14 (1.11–1.18)– El Niño30.0–34.48.5–15.51.12 (1.08–1.16)0.38 La Niña29.4–33.616.4–20.01.06 (1.03–1.09) < 0.001Precipitation Normal31–110420.3–24.61.06 (1.01–1.12)– El Niño84–124410.3–15.41.07 (0.99–1.15)0.973 La Niña93–103516.6–19.61.13 (1.06–1.20)0.183We fit generalized linear models (GLMs) with interaction terms to obtain incidence ratios across temperature and precipitation values. Treating neutral ENSO periods as the reference, p values for interaction are shown. Temperature GLMs adjusted for precipitation at a concurrent lag, while precipitation models adjusted for minimum temperature at the concurrent lag. All GLMs were adjusted for community ID, year, Oceanic Niño Index, and years since insecticide-treated bednet delivery. All temperature predictor ranges are reported in °C. Precipitation predictor ranges are reported in millimeters. Incidence ranges are reported per 10,000 person-weeks. Incidence ratios for temperature predictors are reported per 1°C increase and per 500 mm increase for precipitation. Table 2Plasmodium vivax malaria incidence ratios for malaria incidence across temperature and precipitation, by ENSO (El Niño-Southern Oscillation) period.ENSO periodPredictor rangeIncidence rangeIncidence ratio (95% CI)Interaction p-valueMinimum temperature Normal19.9–23.717.3–31.71.34 (1.28–1.39)– El Niño21.4–24.04.3–16.71.28 (1.21–1.36)0.269 La Niña20.2–23.414.2–20.21.04 (0.99–1.10) < 0.001Maximum temperature Normal29.5–33.720.4–25.91.14 (1.11–1.18)– El Niño30.0–34.48.5–15.51.12 (1.08–1.16)0.38 La Niña29.4–33.616.4–20.01.06 (1.03–1.09) < 0.001Precipitation Normal31–110420.3–24.61.06 (1.01–1.12)– El Niño84–124410.3–15.41.07 (0.99–1.15)0.973 La Niña93–103516.6–19.61.13 (1.06–1.20)0.183We fit generalized linear models (GLMs) with interaction terms to obtain incidence ratios across temperature and precipitation values. Treating neutral ENSO periods as the reference, p values for interaction are shown. Temperature GLMs adjusted for precipitation at a concurrent lag, while precipitation models adjusted for minimum temperature at the concurrent lag. All GLMs were adjusted for community ID, year, Oceanic Niño Index, and years since insecticide-treated bednet delivery. All temperature predictor ranges are reported in °C. Precipitation predictor ranges are reported in millimeters. Incidence ranges are reported per 10,000 person-weeks. Incidence ratios for temperature predictors are reported per 1°C increase and per 500 mm increase for precipitation. Plasmodium vivax malaria incidence ratios for malaria incidence across temperature and precipitation, by ENSO (El Niño-Southern Oscillation) period. We fit generalized linear models (GLMs) with interaction terms to obtain incidence ratios across temperature and precipitation values. Treating neutral ENSO periods as the reference, p values for interaction are shown. Temperature GLMs adjusted for precipitation at a concurrent lag, while precipitation models adjusted for minimum temperature at the concurrent lag. All GLMs were adjusted for community ID, year, Oceanic Niño Index, and years since insecticide-treated bednet delivery. All temperature predictor ranges are reported in °C. Precipitation predictor ranges are reported in millimeters. Incidence ranges are reported per 10,000 person-weeks. Incidence ratios for temperature predictors are reported per 1°C increase and per 500 mm increase for precipitation. Associations between maximum temperature and malaria incidence were positive when stratifying by ENSO period and strongest during neutral and El Niño periods. The incidence ratio for a 1ºC increase in maximum temperature was 1.14 (95% CI 1.11–1.18) during neutral periods, 1.12 (95% CI 1.08–1.16) during El Niño periods (interaction p = 0.38), and 1.06 (95% CI 1.03–1.09) during La Niña periods (interaction p < 0.001) (Table 2). Associations between total precipitation and malaria incidence were positive when stratifying by ENSO period and somewhat stronger during La Niña periods. The incidence ratio for a 500 mm increase in precipitation was 1.06 (95% CI 1.01–1.12) during neutral periods, 1.07 (95% CI 0.99–1.15) during El Niño periods (interaction p = 0.973), and 1.13 (95% CI 1.06–1.20) during La Niña periods (interaction p = 0.183) (Table 2). Discussion Higher temperatures were associated with modestly higher Pv malaria incidence, and total precipitation was associated with moderately higher incidence in an elimination setting in Peru. Associations generally lasted several weeks, with more sustained, stronger associations for precipitation. We only observed associations between malaria incidence and minimum temperatures above 22°C, suggesting that cooler minimum temperatures may be a limiting factor for Pv malaria incidence. Incidence was lowest during the dry El Niño periods and higher in wetter La Niña and neutral periods. For temperature predictors, associations were stronger during neutral and El Niño periods while for precipitation, associations were somewhat stronger during La Niña periods. Overall, periods of higher temperatures and cumulative rainfall were associated with increased malaria incidence in the subsequent several weeks. Our findings suggest that weather influences malaria burden over long, 1–4 month lags in this region. The length of this association likely reflects the impact of weather on vector population dynamics and biting behavior to initate multiple transmission cycles, as well as the potentially overlooked relapse periodicity of Pv malaria. The efficacy of standard of care anti-relapse treatment is approximately 75–80% [43], but effectiveness can be significantly lower due to challenges with adherence [44, 45]. Further, many infections go undetected and untreated due to limited sensitivity of diagnostics, healthcare access challenges, or minimal symptoms which precludes health seeking. As such, the vast majority of Pv infections are estimated to be relapse infections [46, 47], and incidence measured in this study included both new and relapse infections. Additionally, relapses from the tropical strains of Pv are estimated to range anywhere from 2 weeks [48] to 2–9 months, [9] supporting the duration of our lagged findings. Precipitation We found that higher precipitation was significantly associated with higher malaria incidence at lags of 5–16 weeks. Mechanistically, the lasting influence of precipitation in our findings likely reflects increased breeding ground following rainfall, which could impact multiple transmission cycles. The 5–16 week lag may also include the transition from the dry season, when the vector can still thrive in moist environments. Our findings echo the substantial heterogeneity in associations between precipitation and malaria across the Amazon region. i: precipitation > 4.46 cm was associated with higher malaria incidence in the Brazilian Amazon [17]; precipitation was negatively associated with Pv at a 6-month lag in northern Amazon border regions [49]; and precipitation was only associated with increased Pf incidence, not Pv, in the Ecuadorian Amazon [50]. The substantial heterogeneity in associations just in the Amazonian literature likely reflect the strong microclimatic heterogeneity in the region, as well as genetic diversity and adaptation in vector populations. These findings collectively suggest that dynamics between malaria and precipitation cannot be generalized across Amazonian subregions. Temperature We found that higher minimum temperatures and higher maximum temperatures were both associated with higher malaria incidence from 4 to 11 week lags and 4 to 14 week lags, respectively. The results for minimum temperature were largely in accordance with prior literature, while the results for maximum temperature were surprising. As with precipitation, the mixed support for our findings may reflect the substantial genetic diversity, geographic heterogeneity, and climatic variability in the Amazon region [18–20, 22, 51, 52]. Regarding vector ecology, though the thermal transmission optimum of the Pv–Ny. darlingi coupling has not been studied directly, prior studies suggest that Pf and Pv malaria transmission peaks at 25°C and is limited above 30°C [13]. One study found the thermal optimum for malaria vectors is 25°C with an upper limit of 33°C [12], while another found that Ny. darlingi adult lifespan and body size declined and development rate increased as temperatures increased from 20 to 28°C. Therefore, the vector ecology literature supports our finding of positive associations with minimum temperature, which ranged from 20 to 24°C, but contrasts with our findings of positive associations with maximum temperatures, which ranged from 30 to 34°C. One possible explanation for the conflict with maximum temperature is vector biting behavior: Ny. darlingi typically bites humans at dusk or overnight [53, 54], and the highest temperatures occur in the afternoon. However, this explanation does not account for the expected declines in vector populations above 30°C, driven by the negative impacts on larval development, reproduction, and adult longevity above such temperatures [12, 55]. An alternative explanation is that microclimates, topographical variation, and forest cover may moderate higher temperatures in our study site. Given that we used data with 11 km resolution, our findings could not account for such fine scale phenomena. Finally, the vector in our study sites may be adapted to higher maximum temperatures [56] to prevent population decline, given strong microclimate heterogeneity [20] in the Amazon. Our findings also have mixed support in the empirical Pv literature in similar settings. A study of Amazon border regions found that Pv incidence was negatively associated with minimum temperatures between 17 and 25°C at 0- and 6-month lags, in contrast to our findings for minimum temperature, but positively associated with maximum temperatures between 26.8 and 35.2°C 1- and 2-month lags [49], in accordance with our findings for maximum temperature. A study in neighboring Brazil found that mean temperatures between 25 and 31°C were associated with higher malaria risk at 1-week lag but that mean temperatures above 25°C were associated with lower malaria risk [17] at 2- and 3-week lags. However, the study in Brazil used mean temperature, thereby preventing comparisons with our daily extreme measures, and both studies used different modeling structures and lookback periods, preventing direct comparisons. Given that the majority of incident Pv is thought to be relapse [47], one additional potential explanation for the contrasting associations with temperature may be related to the unexplored relationship between environmental triggers and Pv relapse immunology. Though precise evidence is limited, extrinsic triggers of Pv relapse may include co-infection with P. falciparum [57] or other febrile illnesses [4] that result in host inflammation, subsequent primary Pv infections, seasonal changes in sunlight and temperature, and mosquito bites and their associated immune responses [14]. Pv relapse is common following a Pf infection [4], but Pf infections are uncommon in our study site. Dengue, however, is increasingly common [58], and its transmission is positively associated with maximum temperatures similar to those in our study [59–61] and has a higher thermal optimum (29°C) and upper limit (34.5°C) than malaria [12]. Co-infections with dengue could thus occur at higher temperatures and trigger Pv relapses via a similar mechanism that Pf infections are thought to activate. It is also possible that higher temperatures lead to host inflammation and heat stress that activate the hypnozoite. ENSO sub-group analyses We found that incidence was generally lowest during El Niño periods compared to neutral or La Niña periods, when incidence was similar. There is mixed support for this finding in the literature. One study in neighboring Colombia found that caseloads were higher during El Niño events [62], and a more recent study in the Brazilian Amazon also found that incidence was lower during El Niño and La Niña periods [17], contradicting our results. These discrepancies may be explained by differential impacts of El Niño on vector capacity in Amazonian subregions. Variation in vector populations across Amazonian subregions [22, 53, 63] may further contribute to these discrepancies, as climatic shifts can have differential effects on vector survival and biting depending on local species composition and adaptations. Geographic variability also contributes to vector capacity: in Colombia and Brazil, higher temperatures and lower rainfall may lead to decreased river discharge and water stagnation, potentially creating optimal malaria breeding grounds; in contrast, the small, shallow rivers in our study site can dry up completely during strong El Niño periods, such as in August 2023. Additionally, the Amazon has been undergoing historic drought due to global climate change [27], with 2023 being a particularly pronounced drought year. Climate change may be dramatically altering the effects of weather independently and on a longer time scale than any individual ENSO period studied here. Strong drought brought on by El Niño and climate change may have greatly reduced the availability of stagnant pools of water necessary for mosquito breeding, explaining the overall lowest incidence during El Niño periods during our study period. Additionally, temperatures were highest during El Niño periods, potentially limiting mosquito activity. On the other hand, because La Niña periods are associated with wetter conditions [64], in the context of historic drought that has dried up rivers in our study region, increases in rainfall could potentially create temporary pools of stagnant water, explaining the higher incidence during these periods in our study region. We also found evidence of interaction between La Niña periods and weather predictors; associations between malaria incidence and precipitation were strongest during La Niña periods, while associations between malaria incidence and temperature were weakest during La Niña periods. These interaction effects suggest that variation in precipitation drives malaria transmission during La Niña periods, while during neutral and El Niño periods, temperature largely drives transmission. Community type sub-group analysis Incidence was higher in river- versus highway-adjacent communities. In our study area, highway communities are more urbanized with less vegetation, while riverine communities are more vegetated. The combination of water presence with more vegetation in riverine communities may provides more mosquito habitat, thus contributing to higher incidence in riverine versus the drier, more urbanized, and less vegetated highway communities. In sub-group analyses, the association with maximum temperature was stronger in riverine communities. On the other hand, the association with precipitation was stronger in highway-adjacent communities. It is possible that rainfall is more likely to create stagnant pools of water in highway-adjacent communities than riverine communities, where water bodies are larger and rainfall may result in a more modest increase in ideal water bodies for mosquito breeding. Conversely, in river-adjacent communities where mosquitoes can breed regardless of precipitation, temperature may be more limiting, explaining the stronger association with maximum temperature in these communities. The vegetation in riverine communities may also provide cooler microclimates that approximate the malaria thermal optimum better than more urbanized highway communities, which may be hotter. Limitations Our study had several limitations. We were unable to distinguish between initial and relapse infections in our incidence data, a limitation shared with prior observational studies of Pv [65, 66]. We would expect that weather would influence initial and relapse cases over different lag periods, but our analysis was not able to investigate this. A follow-up study using genomic methods or a longitudinal cohort study with temporally dense testing to differentiate between initial and relapse infections could elucidate if weather differentially influences primary versus relapse infections. We were also unable to distinguish which cases were identified via passive versus active case detection, which were done as malaria control efforts by the Peruvian Ministry of Health, greatly affecting case detection and potentially creating bias in our results. However, these interventions were rare and generally done as a reaction to higher passively detected caseloads, suggesting that passive surveillance-detected incidence was already higher when active case detection was done. The study period also included 2 years (2021 and early 2022) where control efforts for the COVID-19 pandemic likely limited malaria transmission as well. While we selected ERA-5 Land remote sensing data for its temporal coverage and lack of missingness, its 11 km spatial resolution prevented identification of small-scale microclimates, such as those created by forest cover and topographical features, and thus did not capture small-resolution variation that may have a large effect on mosquito breeding and survival habitat [67]. Fine-resolution drone imagery [67], as used in other studies, may help overcome this limitation. While we initially considered land cover variables, they were limited in statistical power. Additionally, publicly available surface water data did not reflect our ground observations of surface water in the study site, so we did not include it in this analysis. Further research that considers the influence of non-weather environmental covariates could help elucidate the complicated, heterogeneous dynamics of malaria transmission in this region, and consider mediation under changing climatic conditions. This study was also correlational, and inferred relationships with individual weather variables may be confounded by collinearity and correlation with other weather variables and thus made it difficult to tease out individual direct effects. Conclusions In our study of malaria incidence in Amazonian Peru, we observed generally positive associations with higher temperatures and higher rainfall for extended lag periods beginning 4–5 weeks after an initial weather event and enduring for 1–4 months. We also found lower incidence during El Niño years and higher incidence in riverine communities, which may be used to concentrate resources and time malaria interventions to achieve elimination in this setting. Our findings indicated that malaria burden in this region may be positively associated with higher temperatures. As the successes or failures of malaria interventions may be in part due to weather initiating long-lasting Pv malaria cycles, these findings provide critical context to ongoing malaria elimination efforts in the Amazon region. Supplementary Information Supplementary material 1: Table S1. Deviations from pre-analysis plan. Supplementary material 1: Table S1. Deviations from pre-analysis plan.