Abstract
Climate drives population dynamics, but when the underlying mechanisms are unresolved, studies can lead to seemingly contradictory effects of climate on natural populations. Climate-sensitive vector-borne diseases such as dengue, chikungunya, and Zika are one example where climate appears to have opposing effects in different contexts. In this study, we use a mathematical model to directly connect climate-driven mosquito physiology measured in laboratory studies to observed vector and disease dynamics in the field across ecologically and culturally distinct settings in Ecuador and Kenya. We show that temperature, rainfall, and humidity predict Aedes aepgyti abundances and laboratory-confirmed arboviral incidence across ecologically distinct settings. Further, this trait-based approach resolves seemingly contradictory results from prior field studies and highlights climate conditions where mechanisms remain unresolved. Using this mechanistic model, we tested several intervention strategies and found that reducing immature mosquito habitat or contact rate between mosquitoes and humans are more effective interventions than killing adult mosquitoes. These results can help guide intervention efforts and improve climate change predictions for vector-borne diseases.
Introduction
Climate is a major driver of species interactions and population dynamics, but the mechanisms underlying these relationships are often poorly understood and rarely tested in the field [1]. One of the primary ways that climate impacts populations is through its effects on species’ vital rates [2]. However, these mechanistic effects can lead to seemingly contradictory results in the field because multiple climate variables may act synergistically, with each climate variable potentially affecting multiple vital rates, and their impacts may be nonlinear, changing direction and relative importance across a gradient of conditions. Vector-borne diseases provide an interesting case study to test whether climate sensitive traits measured in controlled, laboratory settings can reconcile seemingly contradictory results from field studies. For example, mosquito-borne arboviral diseases such as dengue, chikungunya, and Zika are clearly climate-sensitive: a body of field research has consistently identified temperature, rainfall, and humidity as important predictors of disease, but sometimes with opposite conclusions about the magnitude and direction of effects of climate on mosquito and disease dynamics [3–8]. For example, dengue incidence correlated with temperature positively in Mexico [9] but negatively in Thailand [10]. We hypothesize that such opposing effects could be simultaneously correct if disease dynamics are context-dependent or nonlinear, and each model describes true disease dynamics but only within a small subset of conditions (e.g., specific locations or seasons).
Understanding the mechanisms that drive disease dynamics is particularly important for arboviruses like dengue, chikungunya, and Zika because they are a major public health burden, vector control is the main method for breaking transmission cycles, and the burden and distribution of these diseases are projected to shift geographically in the future [11–13]. Half of the world’s population is currently at risk of contracting dengue [14]. With no widely available vaccine, vector control remains the primary method for preventing arboviral disease transmission. Existing vector control methods focus on reducing immature habitat, reducing adult populations, or employing personal protection to reduce contact between infected mosquitoes and people [15]. Like other vector-borne diseases with complex transmission dynamics, model simulations can help guide effective intervention efforts [16, 17]. Further, mechanistic models are better suited to predict how climate change will impact future disease burden and distribution, as projected climate conditions are outside the current arboviral climate niche space.
Dengue, chikungunya, and Zika are climate-sensitive diseases because of the ecology of Aedes aegypti, the primary disease vector. Ae. aegypti are anthropophilic, globally distributed mosquitoes that breed in artificial containers with standing water [18, 19]. All mosquito and parasite traits that are important for transmission and linked to metabolism, such as reproduction, development, survival, biting rate, and extrinsic incubation period, are temperature dependent with a thermal optima [20–22]. Humidity is positively associated with mosquito survival because the high surface area to volume ratio of mosquitoes exposes them to desiccation [23, 24]. Standing water from rainfall provides essential larval and pupal habitat for mosquitoes, but the relationship is complex because heavy rainfall can flush away breeding habitats [25–27] and water storage practices during droughts can increase water availability, mosquito abundance, and contact between mosquitoes and people [28–30].
In this study, our goal was to test the extent to which climate-driven mosquito traits drive disease dynamics across two geographically distinct regions and to characterize the effectiveness of different intervention strategies in those regions. Specifically, we asked: 1) how accurately do mechanistic model predictions reproduce observed mosquito and disease dynamics in the field, 2) are there conditions where the model systematically fails to reproduce observed disease dynamics, and 3) what is the relative effectiveness of different intervention strategies given different levels of intervention effort? To answer these questions, we adapted a mechanistic model for arboviral transmission as a function of climate and independently validated the models with data collected on Ae. aegypti abundances and laboratory-confirmed dengue, chikungunya, and Zika cases from two equatorial countries with distinct socioeconomic, geographic, cultural, and disease transmission settings: Ecuador and Kenya (Fig. 1, Table S1). The study sites within each country were distributed across a temperature gradient with similar ranges of humidity and rainfall. Previous studies have found that Ae. aegypti and dengue were positively associated with warm and wet conditions in Ecuador and Kenya [31–34], although other Ae. aegpyti-vectored arboviruses in Kenya such as chikungunya have been associated with warm and dry conditions [35]. In addition to similar climate conditions, both countries have hyperendemic transmission of all four dengue serotypes and have recently experienced outbreaks of chikungunya; yet, arboviral transmission dynamics differ in each country. In Ecuador, dengue is a re-emerging disease with large seasonal epidemics that frequently result in severe dengue [31]; by contrast, in Kenya, dengue has low levels of year-round transmission [36] and intermittent self-limiting outbreaks that are often undetected [37]. Further, compared with South America, sub-Saharan Africa lacks severe dengue, perhaps because African strains of Ae. aegpyti have lower susceptibility to all four dengue serotypes [38], and/or because people of African ancestry are less susceptible to severe dengue [39].
Results
Relationship between model predictions and observed disease dynamics
The dynamic susceptible, exposed, infectious – susceptible, exposed, infectious, removed (SEI-SEIR) compartmental model (Fig. 2) parameterized with temperature-, rainfall-, and humidity-dependent mosquito life history traits was strongly associated with mosquito abundances and disease dynamics across sites and through time. Model-predicted mosquito abundances and field-collected observations of mosquito abundances corresponded with each other in 65% of the surveys (sample size N = 277 site-months) (Table 1), based on whether the z-scores of predictions and observations were within one standard deviation of each other. Based on surveys conducted across all vector life stages in Kenya (only adult mosquitoes were collected in the Ecuador surveys), the SEI-SEIR model had similar correspondence with the abundance of adult mosquitoes (60%, N = 217) to pupae (60%, N = 217), late instars (57%, N = 217), early instars (56%, N = 217), and eggs (50%, N = 216), likely because the dynamics were consistent across life stages. Model-predicted disease cases corresponded with laboratory-confirmed arboviral incidence in 83% of the surveys (N = 388 site-months) (Table 1). We used z-scores for comparison because the model predictions represent total population estimates whereas observations come from sub-samples of the mosquito and human population.
We explored three additional aspects of model fit and found that the model predicted the magnitude of observations moderately well and detected trends through time and differences across sites. Model-predicted mosquito abundances were positively correlated with field-collected observations of mosquito abundances (Pearson’s correlation coefficient r = 0.35, sample size N = 277) (Fig. 3) and predictions and observations synchronously increased and decreased through time within sites (Exact two-tailed sign test, p < 0.05), but the annual proportion of observations predicted by the model differed across sites (F(7,24) = 10.75, p < 0.001). Similarly, model-predicted disease cases correlated positively with laboratory-confirmed arboviral incidence (r = 0.19, N = 388) (Fig. 3) and predictions and observations synchronously increased and decreased through time within sites (Exact two-tailed sign test, p < 0.001), but the annual proportion of observations predicted by the model differed across sites (F(7,18) = 358.8, p < 0.001). Chikungunya and Zika incidence were only confirmed in Huaquillas and Machala, Ecuador (Ndengue = 366, Nchikungunya = 35, NZika = 14); in those sites, arboviral incidence for each disease peaked at different times and corresponded better with model predictions than dengue alone (Fig. 4).
We tested three hypothesized functional relationships between rainfall and mosquito carrying capacity in the SEI-SEIR model (Fig. S1) and found that the rainfall function that correlated most strongly with field observations differed by response variable (mosquito abundance and arboviral incidence) and site (Table 2). We used correlation to determine the best rainfall function because correlation is the most sensitive metric for magnitude and the rainfall function in the model affects the magnitude of mosquitoes via carrying capacity. The model with the left-skewed unimodal (Brière) rainfall function (Fig. S1a), which indicates that mosquito abundances increase with increasing rainfall until some threshold where flushing occurs, described observed mosquito and disease dynamics most often (Table 2).
Identifying conditions that systematically lead to divergence between predictions and observations
To determine if climate-based, geographical, and urbanization factors could explain conditions where the models consistently over- or underpredicted mosquito abundances and arboviral cases (Table 1), we used classification and regression tree (CART) models. We found that the model systematically overpredicted mosquito abundances when there was low to moderate rainfall (<55 mm) and moderate to high humidity (>1.2 kPA) (Fig. 5a). The model systematically overpredicted arboviral cases when there was high mean temperature (>29°C) or high minimum temperature (>24°C) (Fig. 5b). We did not find evidence of any conditions that systematically led to underpredicting mosquito abundances or arboviral cases, likely because there were many predicted and observed zeros. The 29°C breakpoint that we identified for arboviral cases aligns with the point at which the model predicts that the relative basic reproductive number (R0) declines (Fig. 6). However, the CART results suggest that temperature-dependent mosquito traits may be more constrained at high temperatures than previously estimated from laboratory studies, potentially because of daily temperature variation. Previous field studies estimating the effects of temperature on dengue transmission further support this finding where, in general, locations with mean temperature below 29°C show a positive relationship with dengue incidence whereas locations with mean temperatures above 29°C show negative relationships (Fig. 6).
Evaluating the effectiveness of different intervention scenarios
We simulated three intervention strategies at three intensity levels and found that reducing immature mosquito habitat or contact rate between mosquitoes and humans are far more effective intervention strategies than reducing adult mosquito abundance (Fig. 7). Even with high intensity intervention efforts that reduce mosquito abundance by 90% (e.g., spraying large amounts of insecticide), the model indicates that we would expect only 11% fewer human disease cases (Fig. 7; approximately 12 disease cases per 100,000 population). By contrast, a 10% reduction in immature mosquito habitat (e.g., removing containers from the environment that create pools of standing water from rain) or contact rate (e.g., using window screens, mosquito repellent, or wearing protective clothing) would decrease disease cases by approximately 16% and 19%, respectively (Fig. 7; approximately 110 and 187 disease cases per 100,000 population, respectively). Higher intensity efforts that reduce immature mosquito habitat or contact rate by 50% or 90% provides even greater protection, resulting in predicted decreases in disease cases by as much as 96% (Fig. 7; approximately 2,087 disease cases per 100,000 population).
Discussion
Directly observing the influence of climate on species interactions and population dynamics is often challenging because of interacting and nonlinear relationships; here, we directly and quantitatively connect laboratory-based climate relationships to observed mosquito and disease dynamics in the field, supporting the mechanistic role of climate in these disease systems. The trait-based modeling approach helps to reconcile some long-standing inconsistencies in the literature on the effects of climate on arboviral transmission dynamics. Temperature, rainfall, and humidity are commonly correlated with arboviral transmission, but with apparently inconsistent conclusions about which climate variables best predict disease, in what direction, and at what time lags [3–8]. For example, some studies indicate that mean temperature best predicts disease [50–54], while others indicate that minimum temperature [32,45,55,56] or maximum temperature [7,57–59] are better predictors. Rainfall metrics associated with arboviruses vary widely as well, from cumulative rainfall [6,42,53,59] to number of rainy days [60, 61] to rainfall rates and thresholds [27, 55], and these relationships are difficult to measure in the lab (but see [25]). Further, time lags between climate conditions and dengue incidence are variable rather than static: for example, as temperature and daily rainfall increase, the time lags associated with arboviral incidence decrease [55]. A trait-based model allows these varying time lags to emerge from the nonlinear dynamics of transmission, rather than assuming static time lags. Our results highlight that we should not expect the same climate conditions and lags to be important in all settings, but that their combined, nonlinear effects can predict disease dynamics across different ecological and socio-economic settings.
Understanding the mechanisms that drive disease dynamics can help address two critically important research priorities: assessing intervention strategies and projecting impacts from climate change on disease dynamics. While phenomenological models often replicate arboviral disease dynamics remarkably well [62], mechanistic models that capture mosquito population dynamics and interactions between mosquitoes and humans will provide more accurate predictions for the effects of different interventions or projected changes in climate. In this study, we assessed intervention efforts and found that efforts to reduce immature mosquito habitat or contact rate between mosquitoes and people should be much more effective than approaches targeted to removing adult mosquitoes. Further, the intervention simulations suggest that even low and moderate intervention intensity (10% and 50% reductions) will result in a large percentage of disease cases averted. These results are promising for supporting integrated disease control efforts for dengue, chikungunya, and Zika. To help policymakers in Kenya interpret how these results can guide local intervention efforts, we created a shiny app based on the SEI-SEIR model (https://jms5151.shinyapps.io/shiny/).
Comparisons between the model predictions and field observations highlighted several knowledge gaps about climate-disease relationships. While the model generally reproduced patterns of field observations of mosquitoes and disease cases (based on correspondence between z-scores) and observations increased and decreased in unison with the model predictions (based on sign tests), the relative magnitudes only aligned moderately well (based on Pearson’s correlations) and there was significant variation across sites (based on Table 1 and ANOVA results) indicating that climate may be a more powerful predictor for differences across a spatial climate gradient (i.e., across sites) than through time within a site, which supports previous findings [63]. Further, we found evidence that rainfall influences transmission dynamics via its effects on mosquito carrying capacity. However, incorporating this effect in a dynamic model requires some knowledge of how humans differentially influence immature mosquito habitat availability across regions. We show support for three hypothesized relationships between rainfall and mosquito carrying capacity in the field, indicating that the relationship between rainfall and immature habitat is highly heterogenous, which has been found in previous research in Ecuador [28] and Kenya [64]. By examining conditions where the SEI-SEIR model systematically under- and overpredicted mosquito abundances and arboviral cases, we identified additional specific climate conditions that warrant further empirical experimentation. In particular, a variety of traits important for transmission are not well understood towards the physiologically relevant limits of temperature [65, 66] and humidity [67].
Future research can build on this study to better predict the location, magnitude, and timing of arboviral outbreaks and to assess additional intervention strategies. This study builds on previous mechanistic and semi-mechanistic models [50,61,68–71] by combining a suite of temperature, rainfall, and humidity dependent trait functions into one epidemiological model. However, there were several factors that we did not include in this study, such as existing vector control programs, gradients in land use and land cover, infrastructure, and preexisting immunity in the population (Table S1). For instance, in Ecuador, factors such as distance to abandoned properties, interruptions in access to piped water, shaded patios, and use of vector control are also known to influence arbovirus transmission [72], whereas in the study sites in Kenya, factors associated with arboviral transmission are less well studied and there are currently no vector control or local arboviral surveillance programs employed. Future studies could further improve the model by incorporating human immune dynamics associated with interactions among different dengue serotypes [73] or cross-reactivity among viral antibodies [74], differential susceptibility across human age classes [75], and heterogeneity in contact rates between mosquitoes and people based on human behavior and movement [50, 76]. This study suggests that climate is a key determinant of disease dynamics via its nonlinear effects on mosquito and pathogen traits, and that those relationships can be used to predict the timing and locations of disease outbreaks and to assess intervention strategies. Such mechanistic, climate-driven models will become increasingly important to support public health efforts in the face of novel climate regimes emerging due to climate change.
Materials and Methods
Climate data
We collected in situ measurements of daily mean temperature, relative humidity, and rainfall at each study site and interpolated missing data where necessary, as described below. We used temperature and humidity measurements from HOBO loggers and rainfall measurements from rain gauges for sites in Kenya. We used temperature, humidity, and rainfall measurements from automatic weather stations operated by the National Institute of Meteorology and Hydrology in Ecuador. For Kenya, we interpolated missing temperature data from NOAA Global Surface Summary of the Day (Table S4, Fig. S2) and interpolated missing rainfall data from NOAA Climate Prediction Center Africa Rainfall Climatology dataset (Table S4, Fig. S3). For Ecuador, we interpolated missing temperature (Table S4, Fig. S2) and rainfall (Table S4, Fig. S3) data using the nearest study site where possible and otherwise based on long term mean values for the corresponding Julian day. To interpolate missing data, we linearly regressed all measurements taken on the same day in two datasets and then used the linear model to interpolate temperature for the site with missing data based on the climate measurement from the secondary source for the date when the data was missing (Fig. S2-3). For rainfall, we first calculated a moving window of 14-day accumulated rainfall (following [77]) for each day before interpolation. For both Kenya and Ecuador, we interpolated missing relative humidity data based on long term mean values for the corresponding Julian day (Table S4). We then calculated the saturation vapor pressure deficit (SVPD) from temperature and humidity to use in the humidity function because previous research suggests SVPD is a more informative measure of the effect of humidity on mosquito survival compared with relative humidity [67]. To calculate SVPD, we first calculated the saturation vapor pressure as: where (T) is temperature in degrees Celsius. We then calculated SVPD (in kilopascals) as where RH is relative humidity. The final dataset had no missing values for temperature (Fig. S4), rainfall (Fig. S5), and humidity (Fig. S6).
Vector surveys
We collected, counted, sexed, and classified mosquitoes by species, and aggregated the data to mean number of Aedes aegypti per house, month, and year to account for differences in survey effort across months and sites. We collected adult mosquitoes using Prokopack aspirators [78]. In Ecuador, we collected mosquitoes from approximately 27 houses per site (range = 3-57 houses across four sites) every one-to-two weeks during three, four-month sampling periods between July 2016 and August 2018 (N = 147 sampling weeks across four sites) to capture different parts of the transmission season.
We aggregated the Ecuador vector data to monthly values (N = 60 site-month observations) to correspond with the temporal resolution of surveys in Kenya. In Kenya, we collected mosquitoes from approximately 20 houses per site (range = 1-47 houses across four sites) every month between January 2014 and October 2018 (N = 217 site-month observations). In Kenya, we also collected pupae, late instars, and early instars from containers with standing water around the home and collected eggs by setting ovitraps for an average of four days in and around each house monthly. We brought pupae, late and early instars, and eggs to the insectary and reared them to adulthood to classify individuals by sex and species.
Arboviral surveys
For Ecuador, we analyzed laboratory-confirmed dengue, chikungunya, and Zika cases provided by the Ministry of Health (MoH) of Ecuador. The MoH collects serum samples from a subset of people with suspected arbovirus infections, and samples are tested at the National Public Health Research Institute by molecular diagnostics (RT-PCR) or antibody tests (IgM ELISA for dengue), depending on the number of days of illness. Results are sent to the MoH Epidemiological Surveillance and Control National Directorate (SIVE Alerta system). Laboratory-confirmed dengue cases were available for all four sites from 2014 to 2018. Laboratory-confirmed chikungunya cases were available for Machala and Huaquillas from 2015 to 2018. Laboratory-confirmed Zika cases were available for Machala from 2016 to 2018.
For Kenya, we used laboratory-confirmed dengue cases aggregated by site and month between 2014 and 2018 collected in a passive surveillance study on childhood febrile illness in Kenya (NIH R01AI102918, PI: ADL). The study population consisted of 7653 children less than 18 years of age with undifferentiated febrile illness. Children with fever enrolled in the study when attending outpatient care in one of the four study sites (Mbaka Oromo Health Centre in Chulaimbo, Obama Children’s Hospital in Kisumu, Msambweni District Hospital in Msambweni, and Ukunda/Diani Health Center in Ukunda). Local health officers collected comprehensive clinical and demographic data and phlebotomy at the initial visit. We tested each child’s blood for dengue viremia by molecular diagnostics (conventional PCR [79] or targeted multiplexed real-time PCR when available [80]), or serologic conversion at a follow up visit (IgG ELISA [81]).
SEI-SEIR model
We adapted an SEI-SEIR model parameterized for dengue transmission in Ae. aegypti mosquitoes [82] (Fig. 2) to simulate mosquito abundance and arboviral cases through time based on daily weather conditions in eight study locations. The model (equations 3-9), created independently from the observed data described above, allows mosquito life history traits and viral development rate to vary with temperature (t) following [82], mosquito carrying capacity to vary with accumulated 14-day rainfall (r) following [77], and mosquito mortality to vary with humidity (i.e., saturation vapor pressure deficit) (h) following [67]. where
The mosquito population (Nm) was separated into susceptible (Sm), exposed (Em), and infectious (Im) compartments and the human population (Nh) was separated into susceptible (Sh), exposed (Eh), infectious (Ih), and recovered (Rh) compartments (Fig. 2). Climate-independent model parameters (Table 3) included the intrinsic incubation period (δ), human infectivity period (η), birth rate (br), death rate (dr), and immigration/emigration rate (ie). The temperature-dependent SEI-SEIR model was developed by Huber et al. [82] and allows mosquito life history traits and viral development rates to vary according to thermal response curves fit from data derived in laboratory experiments conducted at constant temperatures (Table 4). The temperature-dependent traits include eggs laid per female per day (epd), the probability of egg-to-adult survival (pea), mosquito development rate (mdr), mosquito mortality rate (lifespan-1; μ), biting rate (a), probability of mosquito infection per bite on an infectious host (pmi), parasite development rate (pdr), and probability of mosquito infectiousness given an infectious bite (b). We modified the mosquito mortality rate equation to vary as a function of temperature and humidity by fitting a spline model based on a pooled survival analysis of Ae. aegypti [67] (Fig. S7): where the rate constant (c), minimum temperature (t0), and maximum temperature (Tm) equal -1.24, 16.63, and 31.85 respectively (Table 4), humidity (H) is the saturation vapor pressure deficit, and y is a scaling factor that we set to 0.005 and 0.01, respectively, to restrict mosquito mortality rates within the range of mortality rates estimated by other studies [20, 67]. The linear humidity function has a steeper slope at lower humidity values (equation 11) compared with higher humidity values (equation 12) based on previous research [67] (Fig. S7).
We modeled mosquito carrying capacity, K, as a modified Arrhenius equation following [82, 83]: with T0 and H0 set to the temperature and humidity where carrying capacity is greatest (29°C and 6 kPA) and the Boltzmann constant, (KB), is 8.617 x 10-5 eV/K. We set the activation energy, EA, as 0.05 based on [82]. Since there were no experimental data from which to derive the functional response of mosquito carrying capacity across a gradient of rainfall values, we tested several functional relationships based on hypothesized biological relationships between freshwater availability and immature mosquito breeding habitat, modeling the effect of rainfall on carrying capacity, f(R), as either: where minimum rainfall (Rmin) equaled 1 mm and maximum rainfall (Rmax) equaled 123 mm based on the high probability of flushing [27]. The quadratic function is similar to the rainfall function found in [27] and the inverse function is based on the rainfall function used in [77]. We used rate constants (c) of 7.86e-5 and -5.99e-3 for the Brière and quadratic functions respectively, based on rate constants for other parameters with similar functional forms (Table 4). We scaled the Brière and quadratic functions by y (0. 268 and 0.045, respectively) so that the maximum carrying capacity was approximately equal across all three functions.
To initiate the model, we used site-specific values for human population size and randomly selected one set of values for all sites for the proportion of mosquitoes and humans in each compartment. For Ecuador, we used population estimates from official population projections produced by Proyección de la Población Ecuatoriana, por años calendario, según cantones 2010-2020 (https://www.ecuadorencifras.gob.ec/proyecciones-poblacionales/) with population sizes of 57,366, 279,887, 13,673, and 25,615 for Huaquillas, Machala, Portovelo, and Zaruma, respectively, based on 2017 projections. For Kenya, we estimated the population sizes served by each outpatient care facility by creating a polygon around all the geolocations of study participants’ homes enrolled at each outpatient care facility and summed population count data from NASA’s Socioeconomic Data and Applications Center Gridded Population of the World v4 (https://doi.org/10.7927/H4JW8BX5) within each polygon using ArcGIS v 10.4.1. We estimated population sizes of 7,304, 547,557, 240,698, and 154,048 for Chulaimbo, Kisumu, Msambweni, and Ukunda respectively. We used the following values as the initial proportion of mosquitoes and humans in each model compartment: Sm = 0.22, Em = 0.29, Im = 0.49, Sh = 0.58, Eh = 0.22, Ih = 0.00, and Rh = 0.20. We determined that the model was invariant to initial proportion values after a short burn-in period (90 days) based on a sensitivity analysis (Fig. S8).
We ran all model simulations using the deSolve package in R statistical software v 3.5.3. Model codes is available at https://github.com/jms5151/SEI-SEIR_Arboviruses.
Model validation
To validate the SEI-SEIR model, we quantified the relationships between predicted and observed mosquitoes and laboratory-confirmed disease cases by comparing z-score values, Pearson’s correlations, sign tests, and Analysis of Variance (ANOVAs). To determine whether there was overall correspondence between model predictions and field-collected observations of Aedes aegypti abundances (N = 277 site-months) and laboratory-confirmed arboviral incidence (N = 388 site-months), we categorized observations of mosquito abundance or disease cases as corresponding to the model predictions if the observation fell within one standard deviation above or below the prediction (using z-scores of observations and predictions), overpredicted if the observations were below one standard deviation below the prediction, and underpredicted if the observations were above one standard deviation above the prediction. To assess the correlation of individual survey points through time within sites, we calculated Pearson’s correlation coefficient, r, between model predictions of observations using the cor function in base R, excluding missing data. To determine whether the model predicted directional trends in the dynamics, we determined whether model predictions and observations increased and decreased in unison by first calculating the number of time points between surveys where predictions and observations of mosquito abundances or disease cases synchronously increased, decreased, or stayed constant between surveys and then used the number of time points in agreement and the total number of time points in a two-tailed exact sign test using the binom.test function in R. To test whether climate effects were more important for determining differences across sites or whether climate was differentially predictive in some sites over others, we calculated the yearly percentage of mosquito and disease case observations predicted by the model and used those site-year values in a one-way ANOVA using the aov function in R.
CART model
To investigate conditions where the model systematically over- or underpredicted mosquito abundances and arboviral cases, we used classification and regression tree (CART) models. For each CART model, we used the three correspondence categories (corresponded, overpredicted, underpredicted) as the response variable and a suite of predictor variables. The predictor variables included site (proxy for socioeconomic status and potential prior exposure to disease), country (proxy for genetic, cultural, healthcare, and infrastructure differences), urban/rural, inland/coastal, and climate conditions in the month prior to each survey, a time interval commonly associated with arboviral transmission [6,42,53]. The climate conditions we investigated in the month prior to each survey were minimum, maximum, mean, and variance of daily temperature and humidity and 14-day cumulative rainfall. We conducted the CART analysis using the rpart package in R.
Comparison of R0 with prior studies
We collected effect sizes of temperature on dengue incidence from 12 peer-reviewed studies from the literature (Table S3). We selected studies with mean temperatures across the predicted temperature range where arboviral transmission can occur. We scaled the coefficient values to visualize the relative effect of temperature across studies given that the original analyses were conducted with different temperature metrics and across different temperature ranges. We provide additional information and sources in Table S3.
Intervention simulations
We simulated different intervention strategies by adapting the SEI-SEIR model and simulating disease cases over a one-year time period. We simulated three intervention strategies (reducing contact rate between mosquitoes and humans, reducing immature mosquito habitat, and reducing mosquito abundance) at three intensity levels (10%, 50%, and 90% reduction). Each of these simulation strategies preserves the temperature-, rainfall-, and humidity-dependence of each parameter but modifies the magnitude of one or more parameters. To simulate a reduction in contact rate, we multiplied the mosquito biting rate, a, by 0.10, 0.50, or 0.90. To simulate a reduction in immature mosquito habitat, we multiplied the carrying capacity function equation, K (equation 12), by 0.10, 0.50, or 0.90. To simulate a reduction in mosquito abundance, we reduced the proportion of mosquitoes in the susceptible, exposed, and infectious compartments by 0.10, 0.50, or 0.90. In contrast to the first two interventions that are considered relatively “static” (e.g., adding screens to windows will consistently reduce contact rate), the third intervention represents an activity that is labor intensive and is applied at a single time point (e.g., spraying insecticide). Therefore, for the third intervention, we ran simulations where the intervention occurred once a year and we varied the timing of the intervention by month (e.g., 12 simulations per intensity level).
Author contributions
EAM, ADL, EFL, and JMC conceived of project. JMC conducted analyses and wrote manuscript. EAM, ADL, EFL, and AMSI secured funding for the project. BNN, FMM, EBA, AA, MJBC, RD, FHH, RM, and HNN collected data. ENGS and MMS conducted laboratory analyses. ARK, SJR, and RS processed data. All authors revised and approved of the manuscript.
Acknowledgements
JMC, ADL, EFL, and EAM were supported by a Stanford Woods Institute for the Environment – Environmental Ventures Program grant (PIs: EAM, ADL, and EFL). EAM was also supported by a Hellman Faculty Fellowship and a Terman Award. ADL, BAN, FMM, ENGS, MSS, ARK, RD, AA, and HNN were supported by a National Institutes of Health R01 grant (AI102918; PI: ADL). EAM, AMSI, and SJR were supported by a National Science Foundation Ecology and Evolution of Infectious Diseases grant (DEB-1518681) and AMSI and SJR were also supported by an NSF DEB RAPID grant (1641145). EAM was supported by a National Institute of General Medical Sciences Maximizing Investigators’ Research Award grant (1R35GM133439-01).
References
- 1.↵
- 2.↵
- 3.↵
- 4.
- 5.
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.
- 27.↵
- 28.↵
- 29.
- 30.↵
- 31.↵
- 32.↵
- 33.
- 34.↵
- 35.↵
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.
- 42.↵
- 43.
- 44.
- 45.↵
- 46.
- 47.
- 48.
- 49.↵
- 50.↵
- 51.
- 52.
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.
- 70.
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵